Efficiency Radar

★ TriAttention: Unified Sparse Mixed-Precision Attention for Long-Context LLM Inference

Hanzhi Zhang, Qiao Zhang, Qinglei Cao, Heng Fan, Yan Huang, Kewei Sha — Fri, 06 Nov 2026 00:00:00 +0000

97% · Inference speed 97%, Compression 91%, Serving systems 65%

Sparsity and quantization reduce the cost of attention by skipping interactions and lowering the precision of those retained. However, useful interactions differ in their precision requirements: some tolerate low-precision scores while remaining sensitive to removal. We therefore introduce TriAttention, which assigns query-key (QK) tiles to skipping, 4-bit, or 8-bit computation, using 4-bit to retain quantization-tolerant interactions at lower cost. To represent compactly, we encode each head's computation regions with two diagonal-distance boundaries, keeping policy storage independent of context length. Offline calibration then evaluates candidate boundary pairs using normalized projected-output errors and complete-tile execution costs. Using these measurements, a model-wide allocator selects each head's boundaries to minimize calibration error under a shared computation budget. During prefill, the selected boundaries generate precision-grouped traversals for a fused kernel that skips omitted tiles and combines contributions from both QK precisions into one normalized attention output. The quantized keys and values prepared during prefill are then reused in a compressed key-value (KV) cache for decoding, reducing repeated reads of prompt data. Experiments show that TriAttention achieves end-to-end generation speedups of up to 7.88x over the Torch attention and 2.65x over the BF16 baseline. Meanwhile, TriAttention maintains quality comparable to the original models on general-purpose and long-context tasks. These results show how a compact retention-and-precision policy translates local sensitivities into attention computation and repeated KV-cache access. The implementation is https://github.com/HanzhiZhang-Ulrica/TriAttention.

★ LowRAM: Low-Rank Keys and Amortised Values for Composable KV Cache Compression at Test-Time

Teresa Delgado, Hugo Hazard, Martin Benfeghoul, Jun Wang, Zafeirios Fountas — Fri, 06 Nov 2026 00:00:00 +0000

96% · Inference speed 96%

Autoregressive decoding relies on the KV cache, whose memory footprint grows linearly with context length and caps batch size and throughput. Existing post-training methods exploit redundancy across tokens, layers, and feature dimensions, or lower numerical precision. We identify a complementary axis: within each attention head, values are largely predictable from their corresponding keys. We exploit this with value cache amortisation, which fits a lightweight per-head predictor at test time and stores residuals only for the tokens it reconstructs worst, giving $4\times$ near-lossless value compression. The predictor can be initialised by meta-learning for substantially reduced prefill overhead, or fitted from scratch in settings that allow a one-off cost, such as multi-query inference. We pair value amortisation with cross-layer low-rank key compression to produce LowRAM, an asymmetric compression pipeline that stays close to lossless across three model families. By reconstructing only a selected subset of KV pairs at each decoding step, LowRAM can keep decoding overhead minimal and achieve up to $5\times$ higher peak decode throughput at $4\times$ compression. LowRAM further composes with quantisation and eviction, reaching $26.9\times$ compression at $94.5\%$ of baseline performance on long contexts. These results show that treating keys and values as structurally distinct enables efficient, aggressive, and composable KV cache compression.

★ TileQuant: Tile as the Precision-Allocation Unit for LLM Mixed-Precision Post-Training Quantization

Hanzhi Zhang, Qiao Zhang, Qinglei Cao, Heng Fan, Yan Huang, Kewei Sha — Fri, 06 Nov 2026 00:00:00 +0000

96% · Compression 96%, Inference speed 92%

Post-training quantization (PTQ) reduces large language model (LLM) deployment costs, but preserving quality requires accounting for uneven weight sensitivity. Uniform PTQ uses one weight bit-width, importance-aware PTQ adjusts scales or reconstructs weights within that width, and column-wise mixed precision retains selected columns at higher precision. However, sensitivity varies within columns, so whole-column retention also protects less sensitive weights, while runtime outlier detection adds decoding overhead. We therefore introduce TileQuant, a training-free framework that assigns precision to hardware-aligned 2D weight tiles. Offline calibration provides input-activation-aware scoring from forward statistics for lower calibration cost and gradient-informed scoring from loss sensitivity for stronger quality preservation. Within each projection's retention budget, the highest-scoring tiles remain at 16-bit and the rest use 8-bit. The resulting map guides reconstruction into compact weight buffers, which the kernel reuses for native integer and 16-bit computation with local accumulation to reduce memory traffic. Experiments show that TileQuant reaches 2.66x the original model's generation throughput. Across language modeling, reasoning, long-context retrieval, and question answering, it preserves quality and surpasses quantized and original models on multiple tasks. These results support selecting precision units that capture local sensitivity while enabling efficient hardware execution. The implementation is: https://github.com/HanzhiZhang-Ulrica/TileQuant

★ $D^2$-OPD: Separating Direction and Distance in Extrapolative On-Policy Distillation

Fri, 06 Nov 2026 00:00:00 +0000

96% · Compression 96%

On-policy distillation (OPD) trains a student model to match a teacher's next-token distributions on student-generated prefixes. Extrapolative OPD further extends the teacher's log-probability shift relative to a reference policy to construct stronger learning targets. However, existing methods couple extrapolation direction and distance, allowing the magnitude of the teacher's shift to dictate how far the target moves. We propose $D^2$-OPD, decoupling these two factors: a per-prefix hard KL constraint bounds the target's distance from the teacher, while an adaptive combination of the teacher's post-training shift and the initial student--teacher gap determines the extrapolation direction. We optimize the student toward this target using forward KL, providing learning signals even for low-probability tokens. We evaluate three settings: where standard OPD performs well, where it struggles with base-model initialization, and where no teacher reference is available. On mathematical reasoning benchmarks, $D^2$-OPD attains the highest score in all three settings, outperforming the strongest baseline by up to 9.45 points. Ablations validate the contributions of each component.

★ CAMP: Context-Aware Mapping and Placement for LLMs on 3D-IMC Arrays

Fri, 06 Nov 2026 00:00:00 +0000

95% · Inference speed 95%

Large Language Models (LLMs) incur severe memory-access overheads due to the cost of loading large weight matrices and key-value (KV) caches. Analog In-Memory Computing (AIMC) is an emerging solution that enables computation directly on static weights stored in memory, substantially reducing energy and latency. 3D-IMC further stacks memory arrays into vertical tiers, increasing capacity but revealing a new challenge: how weights are placed in memory determines which matrix-vector multiplications (MVMs) can execute concurrently, affecting latency and throughput. The concurrency aspect of 3D-IMC weight-mapping remains underexplored in the literature, leading to sub-optimal utilization of compute resources on 3D-IMC chips. In this context we propose CAMP (Context-Aware Mapping and Placement), an inference-context-aware weight mapping algorithm for 3D-IMC-based LLM accelerators. CAMP captures the potential resource contention resulting from the architecture, targeted network, and runtime context in a conflict matrix, enabling throughput-optimized weight mapping. This mapping is refined through a profiling-driven feedback loop that efficiently prunes the search space and returns optimized mappings within practical runtimes. CAMP is an offline, software-based solution requiring no hardware changes and improving 3D-IMC compute utilization from 42.4% to 60.3%, delivering up to 2.0x higher 3D-IMC throughput and 1.74x higher end-to-end throughput over state-of-the-art mapping algorithms.

★ Reinforcing What the Student Has Learned: Teacher-Guided On-Policy Distillation

Bozhen Fang, Linhao Zhong, Xuehao Wang, Huanyi Zheng, Hao Chen, Chunhua Shen — Fri, 06 Nov 2026 00:00:00 +0000

95% · Compression 95%

On-policy distillation (OPD) improves language-model reasoning through dense teacher feedback on student-generated responses.We first observe an amplitude disparity in this feedback: on math reasoning responses, style-related positions exhibit 3.5 times the mean absolute teacher--student log-probability difference of mathematical positions.Scaling OPD updates by these discrepancies thus directs much of the update strength toward matching the teacher's expression. We instead take the student's own pretrained base as a reference: the student--base difference records the token preferences acquired during post-training and provides supervision that does not depend on the student's distance from the teacher. Building on this reference, we propose Student-Delta On-Policy Distillation (SD-OPD), which uses centered student--base differences as update coefficients, keeps each coefficient only when its sign agrees with centered teacher--student feedback, and recomputes the student signal every training round. Across 5 teacher--student configurations with Qwen3 students of 1.7B to 8B parameters, SD-OPD improves six-task mathematics average pass@1 (avg@16) over OPD by 7.76 to 10.37 percentage points. Code is available at https://github.com/aim-uofa/sd-opd.

★ CoRD: Coupled Rollouts Fill the Blind Spot of On-Policy Distillation

Guhan Chen, Bohan Li, Jun Rao, Hejin Wang, Songtao Tian, Yexin Xie — Fri, 06 Nov 2026 00:00:00 +0000

95% · Compression 95%

On-policy distillation trains a student on its own rollouts, so states the teacher would reach but the student does not are seldom supervised. We introduce CoRD (coupled-rollout distillation), which keeps the on-policy loss and moves only the distribution of training states toward the teacher, drawing rollouts from a geometric interpolation of student and teacher through the maximal coupling of speculative decoding. The coupling keeps the student's token wherever the two distributions agree and otherwise substitutes a token from the part of the target the student underweights, so interventions are as rare as any exact sampler allows and recede as the student learns. Across six base students from two model families (Qwen3 and Qwen3.5, 0.6B to 4B), CoRD raises average pass@1 over on-policy distillation by 0.9 to 2.0 points and pass@8 by 1.2 to 3.4 points on five public math benchmarks at equal steps and data. On frozen teacher-reached states, a 4B student trained with CoRD reaches in 16 steps a gap closure that on-policy distillation does not reach in 80, and substitution rules intervening at least as often in other directions fall short of CoRD.

QRAE: Query-Relevant Attention Entropy for Context Compression in DeepSearch Agents

Yingjie Tan, Haiquan Qiu, Quanming Yao, Manyi Zhong, Xianzhi Yu, Yaqing Wang — Fri, 06 Nov 2026 00:00:00 +0000

93% · Inference speed 93%

DeepSearch LLM agents repeatedly search and carry retrieved sources into subsequent reasoning steps. This expanding context increases token consumption and latency, although many results do not contribute to later reasoning. We propose Query-Relevant Attention Entropy (QRAE), an online context compression method that uses sparse evidence-assessment heads to decide which intact retrieved chunks should persist across reasoning steps. It scores bounded candidate batches through paired prefills: a query-free control suppresses shared attention responses, while normalized entropy measures the breadth of the positive response within each chunk. High-scoring chunks enter a persistent evidence whitelist and are included in the reasoning context. Uncertain chunks remain outside the reasoning context but are retained in a candidate cache, where they can be retrieved for later reevaluation. QRAE requires neither an additional generative compression pass nor modification of the attention operator. Across multiple DeepSearch benchmarks and reasoning models, QRAE improves task success by an average of 2.35 percentage points and reduces end-to-end latency and reported token usage by an average of 30% and 45%, respectively. These benefits generalize to a model with nearly 100B parameters.

NUMA-Torch: Compiler-Driven NUMA-Aware LLM Prefill on Multi-Socket CPUs

Fri, 06 Nov 2026 00:00:00 +0000

91% · Inference speed 91%, Serving systems 34%

For sufficiently large transformer-based Large Language Models (LLMs), whose parameters cannot be retained in on-chip caches, prefill throughput on multi-socket CPUs is often limited by sustainable DRAM bandwidth rather than peak floating-point throughput. In the standard PyTorch compilation (i.e., TorchInductor) path, model weights are allocated without considering the Non-Uniform Memory Access (NUMA) topology: specifically, the affinity between memory banks and compute cores is disregarded. This causes a substantial fraction of weight accesses to traverse the inter-socket interconnect, severely degrading the effective memory bandwidth available to each socket. We address this bottleneck through a compiler-level solution that makes TorchInductor NUMA-aware. Treating each NUMA socket as an independent tensor-parallel partition, we introduce NUMA-aware Tensor Parallelism (NUMA-TP). This approach transparently shards weights across sockets and schedules compute kernels such that thread pools and memory allocations remain strictly local to their respective sockets using CPU affinity. We evaluate the approach on GPT-2 (117M, 345M, 762M) and BERT (Small, Large), which exercise standard multi-head attention and MLP, as well as Gemma 3 270M, which combines Multi-Query Attention with a GeGLU-gated MLP. By strictly enforcing data locality, our approach successfully mitigates cross-socket traffic, achieving prefill speedups of up to 2.63x on a dual-socket Haswell platform, up to 1.87x on a dual-socket Cascade Lake platform, and up to 2.27x on a four-socket Haswell-EX platform.

Using Transformers for Teaching and Learning via Symbolic Distillation of Neural A-life Agents

Fri, 06 Nov 2026 00:00:00 +0000

91% · Compression 91%

Neural networks controlling behavior of A-life agents can evolve over multiple generations. While effective in making the agents survive, such networks are generally opaque to a human observer. Furthermore, the neural weights cannot be transferred between agents with different neural architectures. Distilling such neural networks into compact symbolic policies can help humans understand agent behavior in A-life. Additionally, it allows the agents to share their knowledge in a compact fashion, improving population-level learning. Distilling a neural policy into a symbolic policy is typically performed by searching the space of symbolic policies to find one that approximates the neural policy. However, even with a domain-specific language the space of policies can be large, necessitating incomplete stochastic search such as genetic algorithms. Yet the distillation remains computationally expensive. We propose to instead use a transformer pre-trained on readily available random percept-action data. Doing so speeds up distillation substantially. We derive the transformer architecture via a neural architecture search. Empirical results compare different distillation methods and the influence of the symbolic space on A-life population dynamics through self-reflection and teaching and learning.

ZYFA: Speed and Accuracy of FP8 FlashAttention on AMD CDNA4

Quentin Anthony, Ganesh Bikshandi, Praneeth Medepalli, Stephen Ebert — Fri, 06 Nov 2026 00:00:00 +0000

90% · Inference speed 90%, Serving systems 82%

Attention is a core building block of frontier models and often dominates compute-bound prefill at long context. FP8 matrix instructions offer twice the peak throughput of BF16, but they accelerate only matrix multiplication, and the softmax between the two matrix multiplications of attention runs on vector units that FP8 does not accelerate. On the AMD Instinct MI355X, the vendor's AITER kernels reach 39% of peak FP8 matrix throughput at 32K tokens, compared with 49% in BF16. Attention is also more numerically sensitive to quantization than feed-forward layers, and unlike NVIDIA's Hopper and Blackwell GPUs, the MI355X does not track the completion of matrix instructions in hardware, so overlapping the softmax with the matrix multiplications depends on instruction order. We present ZYFA, prefill kernels for the MI355X built on three techniques: (1) computing each FP8 probability directly from its score with a multiply-add and an integer conversion instead of an exponential; (2) a ping-pong schedule in which one group of threads computes its softmax while another executes its matrix multiplications; and (3) scaling each block of probabilities up by a power of two before rounding, and normalizing by the sum of the rounded probabilities. On attention inputs captured from three models and timed at sustained clock speeds, ZYFA achieves 1.16-1.76x the throughput of AITER. On the Gaussian inputs of kernel benchmarks at 32K tokens, it achieves 1.26-1.37x at head dimension 128 and 1.99x for DeepSeek-style multi-head latent attention. The sustained clock of the MI355X depends on operand values, so throughput on random inputs does not reliably predict throughput on real model activations. At head dimension 128, ZYFA's numerical error is within a few percent of that of Composable Kernel and Triton and 62-78% lower than that of AITER's. The source code is available at https://github.com/Zyphra/zyfa.

Shallow Encoding Emits Deep Memories

Mingyuan Ma, Yangmin Li, Harris Nover, Zheyu Fu, Po-Han Huang, Alex Settle — Thu, 05 Nov 2026 00:00:00 +0000

94% · Inference speed 94%, Serving systems 65%, Compression 44%

Hybrid language models interleave a few softmax-attention layers with many linear-attention or state-space layers, so every layer keeps a memory of earlier tokens: a key-value cache or a fixed-size recurrent state. Serving them spends more compute and high-bandwidth memory on these memories than the output requires. The prefill runs every layer for every prompt token, and every decode step loads and stores whole recurrent states, which on Qwen3.8-Flash-Next take as many bytes per request as the attention cache of thousands of tokens. We find two properties of trained hybrids that locate this excess. First, writes are easier than reads: writing a deep layer's memory of a token needs little of the deep computation that predicting the next token needs. Trained copies of the deep layers' write paths, which we call emitters, therefore write a prompt's deep memories from the residual stream of a shallower anchor layer, and all prompt tokens but the last run as few as 63% of the layers. Second, recurrent states have a state sink: most of their energy lies in a rank-one term along a fixed direction of each head, and the rest is a low-rank working memory. SEED builds on both: it makes the prefill shallow in depth and the decode state shallow in representation. Each recurrent state is stored as its sink plus low-rank factors, which the Gated DeltaNet, Kimi Delta Attention and Mamba-2 updates extend exactly; we prove that the open-loop error of the periodic compressions back to low rank is bounded independently of their number. SEED keeps the original weights frozen, compresses no attention cache, and runs all layers for every generated token. On six checkpoints of five production hybrids of 30B to over 300B parameters, it is nearly lossless on eleven official exams. Served with SGLang on NVIDIA GB300 GPUs on an agentic-coding trace, it raises the per-GPU throughput of Qwen3.8-Flash-Next 1.2-1.4x.

Give the Drafter a Second Chance to Guess: Parallel Reconsideration for Speculative Decoding

Thu, 05 Nov 2026 00:00:00 +0000

93% · Inference speed 93%

Speculative decoding accelerates large language model inference with a small drafter, whose per-guess acceptance rate determines the achievable speedup. Feature-level drafters share a design choice: every token, easy or hard, is drafted under a single forward pass. Many of the tokens the target rejects are ones the drafter is itself unsure about yet the target answers decisively. To address this, we introduce Parallel Reconsideration: guided by its own uncertainty, the drafter reconsiders its hesitant guesses with extra passes that run in parallel with the one-pass draft, adding negligible latency. A co-designed training recipe and serving-time drafting algorithm append these reconsidered guesses to the original draft tree as refined candidates. Across models and benchmarks, on average, Parallel Reconsideration improves over the baselines by 5.3% to 10.2% on acceptance length and by 3.5% to 9.2% on speedup.

flash-comm: Self-Synchronizing Collective Communication for LLM Decoding

Carrie Wang, Jack Li, Jason Zhao — Thu, 05 Nov 2026 00:00:00 +0000

92% · Inference speed 92%

We present flash-comm, a communication library that achieves efficient transport for LLM decoding by fusing readiness synchronization with data movement. flash-comm introduces self-synchronizing transfers in which each data block carries evidence of its readiness, allowing receivers to directly validate the blocks they need without separate inter-GPU barriers. Based on this principle, a cost-guided protocol selector adapts the readiness encoding to message size and the number of participating GPUs, balancing synchronization overhead against additional data-transfer cost. flash-comm further decouples transport order from reduction order to preserve result consistency across GPUs under asynchronous data arrivals. On a server with eight NVIDIA B300 GPUs, flash-comm reduces all-reduce latency by a median of 28.8% and up to 43.7% compared with custom AR v2 (the state-of-the-art all-reduce implementation). In end-to-end decoding, flash-comm reduces time per output token (TPOT) by 1.7-6.8%. Note that all-reduce accounts for 11.4% of the baseline decoding time, which bounds the latency reduction possible from optimizing this operation alone. The same mechanism also supports standalone all-gather, reducing latency by a median of 23.8% and up to 58.9% across eight B300 GPUs.

vExpert: Virtualizing Expert Storage for Adaptive Load Balancing in Distributed MoE Inference

Wenxun Wang, Xiuhong Li, Yida Wang, Chen Tang, Zongle Huang, Ke Hong — Thu, 05 Nov 2026 00:00:00 +0000

91% · Inference speed 91%, Serving systems 79%

Mixture-of-Experts (MoE) models have become prevailing, yet their distributed deployment suffers from severe dynamic load imbalance due to the conflict between static expert placement and inherent routing dynamism. While host DRAM offers capacity to avoid static device storage for load balancing, we identify existing offloading works fundamentally fail in distributed inference due to two overlooked challenges: (1) coupled performance impacts of runtime Host-to-Device transfers due to load dynamism and shared PCIe topology; (2) system-level memory overhead incurred by DRAM usage at scale. In this paper, we propose vExpert that shifts DRAM offloading paradigm from partial residency to full redundancy through virtualized expert storage. We pioneer the first systematic performance model that jointly captures H2D latency, PCIe contention, and load imbalance, building a lightweight allocator to dynamically map physical experts to virtual slots runtime. A novel disaggregated manager is designed to reduce memory overhead in multi-process environments. Evaluated on DeepSeek-V3.1 within an 8-node H100 cluster, vExpert improves load balance ratio by up to 55% and achieves 1.2$\times$ end-to-end speedup.

★ Multi-layer Indexer: Overlapping Selection and Attention in DeepSeek Sparse Attention

Serdarcan Dilbaz, Arjun Krishna, Kiran Kamble, Daniel M. Bikel — Wed, 04 Nov 2026 00:00:00 +0000

97% · Inference speed 97%, Serving systems 78%

Sparse attention cuts the cost of long-context inference by making the attention matrix sparse: each query attends only a subset of the sequence, chosen by a fixed pattern or, in recent production LLMs, learned per query. In DeepSeek Sparse Attention (DSA), the attention of DeepSeek-V3.2 and the GLM-5 family, every layer is preceded by a lightning indexer that scores the whole cache and keeps the k=2048 highest-scoring tokens for the layer to attend. At long context this indexer scan, not the sparse attention it feeds, is where the decode time goes. But adjacent layers select nearly the same tokens, so a layer can attend the selection of the previous indexer layer instead of its own. This removes the dependency: each indexer then runs concurrently with the attention of its layer. The resulting Multi-layer Indexer (MLI) needs no retraining and no new weights. Its default realization changes no kernel, running inside SGLang's production decode graph; an optional fused decode kernel accelerates it further on DeepSeek-V3.2. MLI composes with IndexCache (the cross-layer reuse GLM-5.2 ships as IndexShare): IndexCache drops the indexer from some layers, and MLI hides the scan on the rest. We serve DeepSeek-V3.2 and GLM-5 under IndexCache's protocol at retentions 1/2 to 1/32 (the fraction of layers that keep an indexer) and contexts 10K to 195K. At equal retention the composition decodes faster than IndexCache: up to 14.5% faster on DeepSeek-V3.2 and 9.8% on GLM-5 at 1/2, and 7.9% and 5.1% at 1/4, largest where IndexCache keeps the most indexers and growing with context. At retention 1/r it is faster still than IndexCache at 1/(2r) while keeping twice the fresh selections: at 128K it reaches 1.32x the unmodified model's decode throughput at 1/4, above IndexCache's 1.28x at 1/8. Its LongBench, RULER and LongBench v2 accuracy stays within a few percentage points of IndexCache's.

★ RAPID: Increasing Efficiency of Internal Distillation via Randomized Methods

Ryan Swift, Konstantinos Psounis — Wed, 04 Nov 2026 00:00:00 +0000

97% · Compression 97%, Inference speed 30%

Knowledge distillation is a technique used to transfer the expertise of large teacher models into small, efficient student models. Recent advances have integrated the distillation of internal activations, more directly training students to learn similar representations as the teacher. One of the core challenges of internal distillation is the dimensional incompatibility between the student and teacher internal representations. Existing methods add learnable matrices to project the larger teacher activations into the smaller representation space of the student. This is a notable point of inefficiency, as learning additional parameters for the projectors increases training costs, only for them to be discarded immediately after training. We introduce RAPID (RAndomized Projection for Internal Distillation) a novel distillation methodology which eliminates the wasteful projection learning by leveraging randomized methods. RAPID utilizes an efficient transform construction based on the Johnson-Lindenstrauss lemma to compress the teacher representations into the student embedding space with strong theoretical guarantees of structural preservation. RAPID also employs a combination of exponential moving average (EMA) and homoscedastic uncertainty weighting in a novel multi-objective function designed for internal distillation. Our experiments in post-training for reasoning demonstrate that RAPID increases training efficiency by as much as 21.1% while maintaining performance comparable to existing methods.

★ P$^2$-MOPD: Accelerating Multi-Teacher On-Policy Distillation via Parameter Priors

Sinuo Wang, Jiacheng Lu, Zeyu Zhang, Maojia Song, Rui Sun, Beidi Luan — Wed, 04 Nov 2026 00:00:00 +0000

96% · Compression 96%

Multi-teacher on-policy distillation consolidates expert capabilities by supervising student-generated trajectories, whereas task arithmetic offers a training-free alternative through linear parameter combinations. We view the linear combination of student and expert parameters as a single differentiable model and propose P$^2$-MOPD, which jointly optimizes the mixing coefficients and the student weights under the same on-policy distillation objective. Unlike task arithmetic, which stays within the span of the frozen student and expert parameters for any choice of merging coefficients, P$^2$-MOPD keeps the student trainable, so the experts act as parameter priors without confining it to this span. Across four domains at 1.7B and 4B scales, P$^2$-MOPD achieves faster and more balanced recovery than MOPD from the same initialization, surpassing its final average recovery in half the training steps. At the end of training, our 1.7B model reaches 94.7% average recovery, versus 80.9% for MOPD.

★ GC-MoE: Overlapping Expert-Parallel Token Dispatch with FP8/FP4 Grouped GEMM via CUDA Green Contexts

Yong Li, Kan Liu, Tao Lan — Wed, 04 Nov 2026 00:00:00 +0000

96% · Inference speed 96%, Serving systems 92%

Under expert parallelism, mixture-of-experts (MoE) layers pay all-to-all token dispatch, grouped GEMMs, and a combine step in every layer. In decode each rank holds only 1-256 tokens: dispatch latency and weight traffic dominate while tensor-core utilization is inherently low. In fused operators such as DeepGEMM's Mega MoE, dispatch and computation overlap only by sharing every SM: dispatch warps compete with TMA/MMA and epilogue warps for registers, shared memory, and issue slots; every phase shares one CTA resource budget, synchronization spans the whole SM array, and every replay ends with a workspace-clean pass and a cross-rank barrier. We present GC-MoE, an open-source FP8/FP4 MoE inference operator for NVIDIA Hopper and Blackwell GPUs that uses CUDA Green Contexts to split the device into a small communication partition and a large compute partition. Each role gets its own kernel, resource budget, and synchronization domain, so dispatch and GEMM tune independently. A three-node CUDA Graph (four on Hopper with fused shared experts) launches the dispatch and grouped-GEMM kernels as concurrent roots; a token-ring handshake starts computation on the first filled ring blocks, and double-buffered receive counters remove the trailing cross-rank barrier between replays. The compute partition runs operand-swapped tcgen05 block-scaled MMA on Blackwell and register-A WGMMA with in-register FP4 decoding on Hopper, with fused SwiGLU, requantization, and optional shared experts. On 8 H20 GPUs GC-MoE accelerates DeepSeek-V4.1-Flash decode by 1.188x on average on both the routed and shared-expert suites (peaks 1.296x and 1.353x); on 8 B300 GPUs it accelerates GLM-5.2 and DeepSeek-V4.1-Flash by 1.106x (1.123x with fused shared experts, peak 1.209x), beating Mega MoE at all 54 measured points at matched accuracy. GC-MoE reserves only 20 of 148 SMs on B300 for dispatch.

★ Coarse Semantic Blocks for Efficient Training-Free Sparse Attention

Sosuke Kobayashi, Shoichiro Yamaguchi, Seiya Tokui — Wed, 04 Nov 2026 00:00:00 +0000

96% · Inference speed 96%

We introduce a training-free sparse attention method that builds a prompt-specific key index to accelerate both prefill and decoding. The method clusters the keys into large semantic blocks that group tokens from noncontiguous positions. Each query position selects blocks by scoring their centroids and attends to the original keys and values within them. Queries that select the same block are processed together, so prefill stays fast even though each query chooses its own blocks. Large blocks make centroid scoring and this grouped processing cheap, so each query can attend to more tokens in the same time. Larger blocks with a larger token budget can thus be both faster and more accurate on long-context tasks than smaller blocks with a smaller budget. We evaluate two hybrid LLMs, Qwen3.5-4B and 35B-A3B, on one NVIDIA H200. Including the cost of building and using the index, whole-model wall-clock speedups over dense attention are 2.3-2.4x for prefill and 1.7-2.0x for decoding at 512K tokens, and 2.9-3.0x for prefill at 1M. The 35B-A3B model stays close to dense attention on five RULER tasks through 1M, on LongBench-v2 overall, and on MRCR through 256K. The 4B model loses accuracy mainly on multi-key retrieval, yet outperforms dense Qwen3.5-2B, whose latency is similar at 512K, on all three benchmarks.

★ LeanSparse: Memory-Efficient Serving for Native Sparse-Attention LLMs

Zaifeng Pan, Chris Wu, Xinwei Qiang, Zhengding Hu, Zhongkai Yu, Yufei Ding — Wed, 04 Nov 2026 00:00:00 +0000

96% · Inference speed 96%, Serving systems 95%

Sparse attention limits the history read by each query, but serving long contexts still requires substantial GPU memory. Existing KV offloading systems retain the full history of indexer keys on GPU, limiting decode concurrency. Prefill's temporary memory demand can further delay admission even when enough memory remains for the same request's subsequent decode. We present LeanSparse, a serving system that reduces both decode's persistent memory use and prefill's temporary memory demand. Our design exploits two observations: prefill consumes historical data one layer at a time, and decode can identify promising candidates before comparing their original indexer scores. During prefill, LeanSparse streams layer histories through two reusable GPU buffers, overlapping next-layer loading with ongoing model computation. During decode, coarse-to-fine indexing searches compact representations and retrieves original indexer keys only for candidate refinement. Separate indexer key and KV caches reuse data across generation steps. We implement LeanSparse in SGLang and evaluate it on DeepSeek-V4-Flash. At 950K context length, LeanSparse supports 2.95$\times$ as many concurrent requests and achieves 2.14$\times$ decode throughput over HiSparse. On multi-turn Open-SWE traces, it reduces median time to first token from 21.6 to 0.6s and median session completion time by 17.7% relative to ECHO.

★ Accelerating Offloading-Based MoE Inference by Interleaving Concurrent Requests

Hongseung Yu, In Gim, Kyunghan Lee, Lin Zhong — Wed, 04 Nov 2026 00:00:00 +0000

96% · Inference speed 96%, Serving systems 86%

Expert offloading enables large mixture-of-experts (MoE) language models to run locally, but conventional batching stalls the GPU while missing experts are fetched. A popular remedy, known as prefetching, is to predict what experts will be needed soon and fetch them using idle link bandwidth. Imprecise prediction, however, wastes bandwidth and GPU memory on fetching unnecessary experts. Our key finding is that prefetching can be harmful when there is an expert cache in GPU memory. This is because as the cache includes more necessary experts, expert prediction becomes less precise. Our key insight is that proper request scheduling beats prefetching. Building on this insight, we present Bevy, an MoE inference system based on runahead inference for concurrent requests. A request whose experts are available can run ahead to later layers while other requests wait for fetches, revealing the experts it needs next. Bevy can then fetch the missing experts so identified while computing ready requests, overlapping computation and fetching without predicting expert selections. Running requests separately, however, repeats reads of shared weights in GPU memory and increases compute time. To limit this cost, Bevy's self-balancing scheduler batches ready requests at each layer and prioritizes lagging requests, adapting their grouping as expert availability changes. Across two local platforms, Bevy delivers up to 3.46$\times$ the decode throughput of llama.cpp and up to 1.55$\times$ that of the fastest prefetching baseline.

★ LoopVQ: Sub 1-bit KV Cache Quantization for Looped Transformers

Minsoo Cheong, Sungjoo Yoo — Wed, 04 Nov 2026 00:00:00 +0000

96% · Compression 96%, Inference speed 95%

Looped transformers are a parameter-efficient alternative to standard transformers for language modeling. By reapplying the same layers across multiple loop steps, recent looped language models match the performance of non-looped models several times their size. While this architecture effectively reduces parameter memory, each loop step still stores its own keys and values, so the key-value (KV) cache grows linearly with the number of loop steps and becomes the main memory bottleneck. We observe that the KV cache of looped models is highly redundant along the loop axis and that the loop steps differ substantially in their sensitivity to quantization. Based on these observations, we propose $\textbf{LoopVQ}$, a KV cache vector quantization method for looped transformers that codes channels of multiple loop steps jointly, and fine-tunes the codebooks to match the output of the full-precision model. At 0.5 bits per element, LoopVQ outperforms existing KV cache vector quantization methods for standard LLMs on general, long chain-of-thought reasoning, and long-context retrieval tasks with Ouro and Huginn models. Its design is orthogonal to these methods, and combined with TaSQ, it retains much of the BF16 accuracy even at 0.25 bits per element. We also implement LoopVQ in SGLang, where it nearly halves the decode latency at a 32K-token context and raises the peak serving throughput by up to $4.3\times$ and $6.1\times$ on the 1.4B and 2.6B Ouro models compared to the BF16 baseline.

★ EpiLoRA: Fusing Many LoRA Adapters into Low-Precision GEMMs for Serving and RL

Hieu Tran, Hong Yu — Wed, 04 Nov 2026 00:00:00 +0000

95% · Inference speed 95%, Serving systems 92%

Low-rank adapters (LoRA) let one base model serve many fine-tuned variants: multi-tenant engines keep a single copy of the base and batch the requests of many tenants, each with its own adapter. As base models move to 8-bit (FP8) and 4-bit formats, their general matrix multiplications (GEMMs) become cheaper, but the adapters do not: engines run them as separate small GPU kernels beside the GEMM, whose latency takes a growing share of every decoding step. An adapter projects its input down to a low rank and back up, and our key observation is that the GEMM needs the down-projection only in its epilogue, the final stage in which it writes its output. EpiLoRA, a plugin for the vLLM engine, therefore launches the down-projection first and the low-precision GEMM as a dependent kernel that streams its weights meanwhile and adds every tenant's up-projection in its epilogue. On NVIDIA Hopper GPUs, EpiLoRA decodes for a single adapter with up to 24% lower time per token than vLLM's fastest configuration, and faster than SGLang and TensorRT-LLM; with many tenants, its adapter kernels deliver more throughput than vLLM's on most GPU-model pairs we tested. The same low-precision base also trains: a LoRA reinforcement-learning trainer that computes on the engine's FP8 weights needs no 16-bit copy, so a 27B model trains on one 48 GB GPU, and an engine that then keeps activations in 16 bits removes most of the remaining mismatch between the model generating training data and the one trained.

Workload-Static Optical Circuit Switching For Scalable LLM Inference Decode

Wed, 04 Nov 2026 00:00:00 +0000

94% · Inference speed 94%

LLM inference decode executes fine-grained collectives whose latency directly limits per-user interactivity, yet scale-up fabrics are still built from electronic packet switches designed for training. We observe that, for a given model deployment and parallelization, the communication structure of decode is fixed, and propose workload-static optical circuit switching (OCS): circuits are configured once per deployment, co-designed with model parallelism and with topology-aware collective schedules that turn spare accelerator ports into fewer communication phases or higher bandwidth. We evaluate this design with GXNET, a graph-based workload emulator coupled to analytical and flit-level network models grounded in hardware measurements. In a 576-accelerator scale-up domain serving 5T-parameter MoE models with up to 5M tokens of context, all-OCS fabrics improve interactivity by up to 65% and energy efficiency by up to 119% over 4:1 oversubscribed packet-switched fabrics.

LoopKV: Loop-Aware KV Cache Compression for Serving Looped Transformers

Tianle Zhou, Yang Zhang, Tianyu Wu, Zhuohan Wang, William Nixon, Juncheng Yang — Wed, 04 Nov 2026 00:00:00 +0000

93% · Inference speed 93%, Compression 81%, Serving systems 76%

Looped transformers apply a shared block several times per token to improve accuracy without adding parameters. For models such as Ouro that keep a separate key-value (KV) cache per loop, cache memory limits the batch size that amortizes repeated weight reads. Uniform-bit quantization and fixed rules per loop do not allocate precision according to each layer's sensitivity in each loop. Measurements on Ouro show that quantization sensitivity varies across (loop, layer) cells, and the sum of individual cell damages approximates the overall output damage. Sharing damage also varies by cell, with later loops generally tolerating sharing better. Skipping low-attention values to save reads produces errors that compound across loops. LoopKV uses a one-time offline profile to choose 4-bit storage, 2-bit storage or reuse of the final loop's KV for each cell. The planner minimizes reads under budgets for predicted damage and storage. During decoding, each loop reads the high 2 bits of each 4-bit value and fetches the low bits only where its own attention weights require 4-bit precision. A single fused kernel in vLLM executes the plan. Serving Ouro-1.4B on one RTX PRO 6000, the memory-optimized LoopKV-M plan stores 35% less KV than uniform 4-bit quantization (KIVI-4) on the same fused kernel and CUDA graph. It sustains 1.7-2.2x KIVI-4's goodput under a latency service-level objective with a 16K-token maximum length. None of its averages on six public benchmarks falls significantly below BF16's.

Beyond Compression Ratio: High-Throughput KV Compression for Disaggregated LLM Serving

Junhong Wu, Zedong Liu, Xinyang Ma, Yuhan Chen, Guangming Tan, Dingwen Tao — Wed, 04 Nov 2026 00:00:00 +0000

92% · Inference speed 92%, Serving systems 66%

KV cache compression reduces the communication volume of disaggregated large language model serving, but its practical benefit depends not only on compression ratio and model quality, but also on how efficiently the compression pipeline executes. This requirement becomes increasingly important as modern serving systems adopt high-bandwidth interconnects, where inefficient compression can offset a substantial portion of the saved communication time. This paper presents \system, a high-throughput KV cache compression data path built on top of \kvserve. \system improves compression execution in two complementary ways. First, it fuses the distribution transformation and low-bit quantization stages into a KV-aware \tilelang kernel, eliminating intermediate tensor materialization and redundant memory accesses. Second, it replaces the original lossless compression backend with an open, extensible, and searchable backend based on \lc, enabling high-performance compression operator pipelines. The fused kernel achieves up to $5.9\times$ speedup, while the new lossless backend increases compression throughput from below 90~GB/s to over 500~GB/s. Across the evaluated workloads, \system reduces job completion time by up to $2.04\times$. End-to-end evaluation shows that these optimizations allow KV compression to retain meaningful communication benefits under modern high-bandwidth network.

★ BitLLM: Efficient Low-Bit, Conversion-Hiding LLM Decoding on Mobile GPUs

Hangyu Zheng, Zhihao Shu, Kunxiong Zhu, Minghai Qin, Xiaolong Ma, Gagan Agrawal — Tue, 03 Nov 2026 00:00:00 +0000

97% · Inference speed 97%, Serving systems 75%

The rapid scaling of large language models (LLMs) has unlocked remarkable capabilities across various domains. However, deploying LLMs on mobile devices remains challenging because of their high computational and memory requirements. Low-bit quantization (e.g., 4-bit weights, 8-bit activations) can reduce model size while preserving accuracy, but its practical efficacy on mobile GPUs has been limited so far. In this paper, we identify that the performance gap stems primarily from frequent data type conversions between quantized and floating-point values, and dynamic and inefficient memory access patterns caused by the growing Key-Value (KV) cache during autoregressive decoding. We introduce BitLLM, a mobile inference engine that addresses type conversion overhead and suboptimal memory accesses in quantized LLMs. Our approach features an adaptive layout and packing strategy for the KV cache that ensures efficient memory access throughout LLM decoding. This is coupled with an integer-native execution pipeline that defers and overlaps type conversion outside the compute hot path, keeping the critical MatMul stages in the integer domain. Implemented on off-the-shelf mobile GPUs, BitLLM outperforms leading engines, i.e., LLAMA.CPP, MNN-LLM, LiteRT-LM, and MLC-LLM, achieving average speedups of 1.78$\times$, 1.73$\times$, 1.58$\times$, and 1.47$\times$, respectively, across various models and quantization levels. Under long-context generation, BitLLM reduces average energy consumption (per token) by up to 88.9%.

★ Order-Only On-Policy Distillation

Guangzeng Han, Yu Wu, Xiaolei Huang — Tue, 03 Nov 2026 00:00:00 +0000

97% · Compression 97%

On-policy distillation (OPD) trains a student on its own generations by matching the teacher's token probabilities. These probabilities entangle two kinds of information: which tokens the teacher prefers, and how confident it is. The latter reflects the teacher's own capacity and calibration, which a smaller student may be unable to reproduce. Because the teacher temperature changes these values without changing the ordering, OPD is also sensitive to the temperature choice. We ask whether the teacher's ordering of candidate tokens is enough, and propose Order-Only On-Policy Distillation (O$^{3}$PD). At each student-generated prefix, O$^{3}$PD trains the student to reproduce the order of the teacher's top-$K$ tokens, without fitting their probability values. The objective is therefore invariant to the teacher temperature and needs no temperature tuning. Across six teacher--student pairs, O$^{3}$PD outperforms OPD on the average over HMMT25, AIME24, and AIME25 in every pair, by up to 2.9 points, with gains of up to 5.6 points on AIME25. Other ranking losses yield similar gains, so the benefit comes from ordinal supervision itself. Analysis shows that O$^{3}$PD departs further from the teacher's probabilities than OPD, yet reproduces its ranking more faithfully and yields a stronger student. The ranking, not the probability values, thus carries the knowledge that on-policy distillation needs to transfer.

★ MXCode: Vector Quantization for MX-Native Inference at Scale

Tue, 03 Nov 2026 00:00:00 +0000

96% · Compression 96%, Inference speed 90%

Memory capacity and bandwidth are major bottlenecks in LLM inference, motivating weight compression beyond conventional 4-bit quantization. Vector quantization (VQ) enables approximately 2-bit weight storage but often relies on codebooks incompatible with native FP4 computation. We introduce MXCode (MX-Native Codebook Quantization), a framework bridging ultra-low-bit VQ compression with MicroXcaling (MX)-native FP4 computation. MXCode proposes three strategies: PostScale, which quantizes learned codebooks to MXFP4; PreScale, which extracts MX scales before codebook learning; and PreScale+, which further compresses these scales using VQ. These strategies explore tradeoffs among quantization quality, scale overhead, and codebook granularity. All reconstruct MX-native W4A4 operands without materializing intermediate high-precision weights. We evaluate MXCode on Qwen3 (235B), DeepSeek-R1 (671B), and GLM-5.2 (753B) using the Microsoft MX Emulation Library. MXCode achieves approximately 2 bits per weight with competitive, workload-dependent model quality. On GB300 hardware, experiments with Qwen3 (235B) show approximately 50% lower model-load memory than Native FP4 while maintaining near-Native-FP4 serving throughput. Under memory-constrained prefix reuse, P90 TTFT improves by up to 94.5%. These results demonstrate software-based reconstruction on existing FP4 hardware and inform the development of future MX-native inference systems.

★ GateTrim: Adaptive-Rank Quantization for Large Language Models

Hongwei Ren, Jincheng Ren, Yuchuan Zhou, Jingchun Ling — Tue, 03 Nov 2026 00:00:00 +0000

96% · Compression 96%, Inference speed 82%

We introduce GateTrim, an adaptive-rank post-training quantization method for large language models that learns matrix-specific ranks through gate optimization and threshold-based rank selection. GateTrim standardizes the low-rank factors of a quantized low-rank--residual representation, learns a separate gate vector for each target weight matrix, and uses a shared threshold to automatically determine its retained rank. The resulting ranks can differ across matrices, and the compact quantized representations are subsequently reconstructed. We evaluate GateTrim under W4A8 on Qwen3-4B/8B/14B, Llama-3.2-3B-Instruct, and Llama-3.1-8B-Instruct. Across all five models, we observe a counterintuitive finding: increasing the threshold from $\tau=0.3$ to $\tau=0.7$ reduces the mean retained rank by 46.60%--67.89% while simultaneously lowering WikiText-2 perplexity by 0.59%--5.17%. Threshold 0.7 achieves the lowest mean retained rank and final perplexity among the three tested thresholds, 0.3, 0.5, and 0.7, and reduces measured target-export size by 1.50%--4.25% relative to 0.3. GateTrim uses an $8\times512$-token fitting pool and a separate development pool of the same size. Its Qwen3-4B/8B/14B perplexities at $\tau=0.7$ are numerically below TwinQuant's published W4A8 values, reported with a $128\times2048$-token calibration pool, in a cross-study comparison. On Qwen3-8B, the threshold-0.7 configuration already has the lowest perplexity before reconstruction under both tested initialization seeds; reconstruction improves it further without increasing retained rank or export size.

★ Jet-QSparse: Accelerating Agentic LLM Inference with 4-Bit Sparse Attention

Haozhan Tang, Hao Kang, Meng Lu, Yecheng Wu, Zerui Wang, Seokho Han — Tue, 03 Nov 2026 00:00:00 +0000

96% · Inference speed 96%, Compression 88%, Serving systems 77%

Long-horizon reasoning and agentic workloads strain KV-cache capacity and decode attention throughput. We present Jet-QSparse, a training-free method that combines compact NVFP4 cache storage with native FP4 query-key computation on Blackwell tensor cores. A dense scan scores the full key history and selects 128-token blocks for sparse FP16 value aggregation with FP32 accumulation, preserving higher precision in the output-forming step while enabling contiguous tile loads. Calibrated key centering and at most two BF16 pages per sequence and layer address quantization error; FP8 prefill uses a temporary arena while the persistent cache is packed directly from BF16 activations. Across four hybrid LLMs, a budget of 128 blocks per KV head shows no detected accuracy loss on the evaluated mathematical reasoning, coding, and agentic benchmarks, although some smaller budgets incur losses. With 64 blocks, Qwen3.8-27B achieves a $1.94\times$ end-to-end decode speedup over dense BF16 at 128K context and batch size 16 on one GB300, and the measured Qwen KV pools provide $2.58$-$2.76\times$ BF16's token capacity.

★ Multi-Turn Semantic Caching for Low-Cost LLM Serving

Baran Atalar, Selly Beyene, Xutong Liu, Carlee Joe-Wong — Tue, 03 Nov 2026 00:00:00 +0000

96% · Inference speed 96%

Semantic caching lowers the cost of serving large language models (LLMs) by reusing stored responses for similar requests. In multi-turn conversations, whether a cached response is acceptable depends on both conversations' histories, which embedding similarity of requests and histories predicts poorly (AUC 0.62--0.74 on five workloads). We therefore learn the \emph{reuse risk}, the probability that a cached response is unacceptable, from judge feedback, and choose a budgeted set of conversation states to cache by weighing the generation cost a reuse avoids against a price for unacceptable reuse. We propose CUCB-SC-MT, which learns a cache from logged conversations with a suboptimality guarantee, and CLCB-LS-MT, which learns online and has $\tilde O(\sqrt N)$ regret over $N$ requests when conversation states recur. On math-tutoring (SocraTeach) and movie-ticketing (Taskmaster-3) conversations, both achieve lower loss than prior semantic caching approaches: ContextCache, multi-turn variants of ElastiCache and vCache, and single-turn semantic caching at every cache budget. With 1,000 cache entries, CLCB-LS-MT avoids 35--41\% of the generation cost through acceptable reuses, against at most 22--25\% for the baselines, and lowers estimated serving latency by 19--35\%.

★ Layer-wise CPU-GPU Scheduling for LLM Inference on Memory-Limited Consumer GPUs

Naoto Sugiura, Yuta Tokusashi, Hiroki Matsutani — Tue, 03 Nov 2026 00:00:00 +0000

96% · Inference speed 96%, Serving systems 47%

As large language model inference environments expand toward consumer devices, it remains challenging to efficiently run models that exceed GPU memory capacity in single-GPU settings. The latency of transferring weights that do not fit in GPU memory accounts for about 98\% of inference time and remains a major obstacle. In this paper, we focus on the fact that during the decode phase, layer computation time on the CPU can be shorter than weight transfer time, presenting an opportunity to hide weight transfers by utilizing this computation time. Therefore, we propose a method to hide transfer latency by executing certain layers on the CPU during the decode phase while asynchronously transferring the next layer's weights to the GPU in the background of that computation. To realize this scheduling, we formulate the placement of CPU/GPU computation and weight transfers as a latency minimization problem and perform optimization using an approximation model with relaxed temporal constraints. Evaluation results confirm that in a consumer desktop environment, our approach accelerates OPT-30B inference by 3.5x compared to HuggingFace Accelerate and 6.3x compared to DeepSpeed Zero-Inference.

★ SpecPilot: Autonomous Algorithm Discovery and Systems Optimization for Speculative Decoding

Kunjun Li, Wenhao Chai, Yihao Liang, Yuezhou Hu, Shuyao Xu, Yi Pan — Tue, 03 Nov 2026 00:00:00 +0000

96% · Inference speed 96%, Serving systems 34%

Speculative decoding accelerates memory-bound LLM decoding with a lightweight drafter, and parallel drafters extend this speedup by drafting a whole block in one forward pass. However, it remains unclear whether their released systems (stock) attain the full speedup that parallel drafting permits. We identify three challenges. First, with one request, each step accepts few tokens but reads every target weight. Second, with batching, per-request verification work roughly halves the speedup over autoregressive (AR) decoding. Third, speculative decoding is deployment-specific, and no universal optimization fits every serving regime, drafter, target model and GPU. To address these challenges, we present SpecPilot, an autonomous agentic loop that discovers algorithms and systems optimizations for each deployment. In each round, a cost model locates the largest gaps between the current system and its ceiling. Parallel agents explore directions and integrate the successful ones. Token forcing checks every output token against stock to detect reward hacking. A structured memory carries validated mechanisms to new rounds and deployments. SpecPilot discovered a general lossless 12-bit weight format, tree verification for hybrid recurrent targets, two-level verification and three batching mechanisms. We evaluate it across four target models, three drafters, two GPU platforms and five benchmarks. Without training or quality loss, it raises the speedup over AR from stock's 4.46$\times$ to 10.87$\times$ with one request, and from 1.94$\times$ to 3.93$\times$ with batching.

★ Rethinking KV Cache Eviction for Agentic LLM Serving

Tue, 03 Nov 2026 00:00:00 +0000

95% · Inference speed 95%, Serving systems 80%

Key-value (KV) caches impose substantial memory pressure on large language model serving systems, requiring cache eviction when capacity is insufficient. Existing mechanisms typically evict KV from the ends of cached sequences to preserve prefix sharing across requests. However, agent workloads repeatedly reuse long histories across successive turns. Recomputing the KV cache for tokens near the end of a sequence is substantially more expensive than for earlier tokens because their attention computation involves longer histories. To address this problem, we propose preserving shared prefixes and evicting the KV cache from the beginning of the remaining history while retaining the cached suffix. We build Penelope around this strategy by co-designing the eviction policy, KV recomputation, and unified GPU-CPU cache management. Returning requests reuse the KV cache retained in GPU or CPU memory and recompute only missing segments while preserving reuse of the cached suffix. We implement Penelope in SGLang and evaluate it on three Qwen3 model sizes using real agent traces. Compared with SGLang, Penelope reduces mean time to first token by up to 60% and job completion time by up to 37% with a GPU-only cache, and by up to 76% and 60% with a two-tier GPU-CPU cache.

★ TileWeave: Heterogeneous Tiling for Sparse Prefill on Mobile SoCs

Wentao Ni, Hong Cai, Mingu Lee, Chihao Yu, Zhongming Yu, Hejia Zhang — Tue, 03 Nov 2026 00:00:00 +0000

95% · Inference speed 95%, Serving systems 61%

Smartphone applications increasingly require large language models to process long contexts, whereas high prefill latency and memory demands constrain local inference. Block-sparse attention reduces computation and memory traffic, but its irregular selections conflict with mobile NPUs' preference for large, regular tiles. Efficient execution must exploit cross-query reuse while avoiding redundant computation and preserving selections needed for accuracy. In this paper, we present TileWeave, which partitions sparse attention by this cross-query reuse: highly shared key blocks run in coarse NPU tiles, while low-reuse blocks run on the GPU only for the query blocks that selected them. In addition, we design an NPU-native pooled-key selector that keeps selection inexpensive, and a runtime that publishes GPU partials in the order the NPU consumes them and merges them in its existing output store. On RULER, HeteroSparse matches XAttention's accuracy with Qwen3-1.7B and improves it with Llama-3.2-3B-Instruct; its mean LongBench scores match XAttention as well. On Snapdragon SM8750 and SM8850 phones, it achieves 1.35--3.23$\times$ faster full-model prefill than XAttention over 1k--16k tokens and up to 1.58$\times$ faster than dense attention.

★ Weights That Stay: On-Chip Megakernels for Serial Heads in Speculative Decoding

Xinwei Qiang, Ziyi Xu, Zaifeng Pan, Chang Chen, Yufei Ding — Tue, 03 Nov 2026 00:00:00 +0000

95% · Inference speed 95%, Serving systems 83%

Parallel speculative drafters use a lightweight serial head to recover token dependencies within a block generated in parallel. Each head step updates only a small state, yet projects that state through the same vocabulary-sized matrix. Step-wise execution repeatedly transfers this large, immutable projection, making the head a meaningful source of latency despite its low arithmetic cost. We present Weights That Stay, an execution design that keeps projection shards on chip for the entire serial draft walk. A persistent megakernel couples weight residency with GPU-side token selection and state advancement, reusing resident weights while exchanging only compact state between steps. The design supports full or partial residency and accommodates both Markov and recurrent heads on H100 and B200. Integrated into SGLang, Weights That Stay improves throughput for Markov drafting on Qwen3-4B and Qwen3-8B, with gains of up to 12.3%. At the head level, the DSpark walk is up to 11.6$\times$ faster than the fastest measured SGLang reference path. These results show that aligning weight residency with the full draft walk translates cross-step reuse into end-to-end serving gains.

★ Search-Aware On-Policy Self-Distillation with Sibling Alignment for Generative Recommendation

Rongfeng Guo, Zhuang Zhuang, Ji Dai, Peng Zhao, Shijie Li, Jie Chen — Tue, 03 Nov 2026 00:00:00 +0000

95% · Compression 95%

Generative recommendation retrieves items by autoregressively generating semantic IDs, typically via beam search. On-policy self-distillation (OPSD) enhances training supervision by transferring privileged self-teacher knowledge at student-generated states. However, beam-search retention depends on competition among prefixes: visited states differ in importance, and initially reliable decisions may approach the pruning boundary as training evolves. Moreover, conventional distribution matching couples target-token confidence with preferences among competing continuations, potentially weakening relational supervision when the self-teacher strongly favors the target. To address these challenges, we propose Search-Aware On-Policy Self-Distillation (SA-OPSD), a framework that transfers privileged self-teacher knowledge to the states visited during the student's own beam search. Specifically, we introduce Evolving-Boundary Sampling, which prioritizes search-critical states by considering both current pruning-boundary proximity and initial-reference retention-margin decline under a controlled distillation budget. Furthermore, we develop Confidence-Decoupled Branch Distillation, which separates target-probability calibration from informative sibling relations to better preserve the teacher's preferences among competing branches. Experiments on Industrial, Office, and Instruments demonstrate strong performance for SA-OPSD.

★ MatMoE: Matryoshka Mixture-of-Experts with Dynamic Mixed-Precision Quantization for Efficient Inference

Renjie Wei, Haochen Huang, Chuyu Qiu, Jinqi Wen, Dehao Xu, Meng Li — Tue, 03 Nov 2026 00:00:00 +0000

95% · Inference speed 95%, Compression 93%, Serving systems 35%

Mixture-of-Experts (MoE) architecture has been widely adopted in large language models (LLMs) since it reduces the computational cost through sparse expert activation. However, the massive parameter scale of MoE models hinders their deployment on resource-constrained edge devices. Although existing mixed-precision quantization reduces the model size, they are inherently static, fixing the bit-width of each expert across decoding steps. Such static schemes yield suboptimal performance as they overlook the dynamic variation in expert importance during decoding. Furthermore, they fail to adapt to the dynamic latency constraint driven by the fluctuating system utilization. To address these challenges, we propose MatMoE, a dynamic mixed-precision quantization framework for efficient MoE inference. MatMoE features: (1) a Matryoshka MoE PTQ algorithm that enables experts to operate flexibly at multiple precisions within a stringent memory budget; (2) a floor-and-compensation quantization method that eliminates redundant memory accesses in the Matryoshka representation; and (3) a dynamic bit-width allocation strategy that determines the precision for each expert based on their routing scores to satisfy the dynamic latency requirement. We implement our MatMoE system atop llama.cpp for both CPU and GPU platforms. Extensive evaluations across three modern MoE models and challenging benchmarks demonstrate that MatMoE achieves $4.0\% \sim 18.3\%$ higher accuracy than state-of-the-art static mixed-precision method. Compared to llama.cpp, MatMoE delivers $1.3\times \sim 3.7\times$ speedup on Raspberry Pi 5, and $1.2\times \sim 1.3\times$ speedup on NVIDIA RTX 4090 and Jetson AGX Orin.

★ OC-OPD: Observation-coordinated Supervision for On-Policy Distillation

Ruoran Xu, Yilun Zhao, Ziyu Chen, Arman Cohan — Tue, 03 Nov 2026 00:00:00 +0000

95% · Compression 95%

Reverse and forward Kullback--Leibler (KL) divergences provide complementary supervision in on-policy distillation: reverse KL refines the student's active token rankings, and forward KL allocates mass to teacher-supported candidates. At student-generated prefixes, we form a four-observation matrix by treating the student and teacher as both candidate sources and evaluators. The diagonal self-observations measure uncertainty, while source-centered cross-observations recover the two KL objectives. We establish an exact identity equating the Fisher inverse-metric coupling of their logit gradients with the sum of these contrasts, yielding a common local descent direction for distinct positive distributions. Observation-coordinated on-policy distillation (OC-OPD) uses observations to set preferences, closed-form bargaining to coordinate gradients, and distributional discrepancy to scale the update. Across three heterogeneous student settings (non-thinking generation, thinking generation, and base-model initialization), OC-OPD achieves strong gains in mathematical reasoning and transfers effectively to out-of-distribution scientific question answering and code generation. Our code is publicly available at: https://github.com/RuoranXu/OC-OPD.

★ Bernoulli On-Policy Distillation: Don't Ignore High-Confidence Student Tokens

Mingyu Cao, Yaokai Jin, Shiwei Liu, Lu Yi — Tue, 03 Nov 2026 00:00:00 +0000

95% · Compression 95%

On-policy distillation (OPD) trains a student language model on its own trajectories using token-level supervision from a stronger teacher. We show that standard sampled reverse-KL OPD systematically suppresses updates to high-confidence tokens, such that their weak gradient contribution does not necessarily imply weak or uninformative teacher supervision. We introduce Bernoulli On-Policy Distillation (Bernoulli OPD), which directly matches the teacher and student probabilities for that token. This yields a bounded probability space correction that preserves teacher-student disagreement while avoiding the additional confidence-dependent suppression induced by vanilla OPD. Fixed-rollout analyses show that Bernoulli OPD substantially restores optimization signal to high-confidence tokens while remaining sensitive to teacher-student disagreement. We further analyze the resulting model behavior through confidence distributions, answer calibration, policy movement, and training dynamics, showing that Bernoulli OPD's gains cannot be explained simply by longer reasoning, larger optimization steps, or more uniform imitation of the teacher. Code is available at https://github.com/duterscmy/Bernoulli-OPD.

★ Accelerating LLM Decoding with In-DRAM Parallelism

Kevin Xu, Tatsuya Kubo, Daichi Tokuda, Shinya Takamaeda-Yamazaki, Issei Sato — Tue, 03 Nov 2026 00:00:00 +0000

95% · Inference speed 95%, Serving systems 37%

Autoregressive decoding of large language models (LLMs) on edge devices is dominated by memory-bound general matrix--vector multiplication (GeMV) operations. Processing-using-DRAM (PUD) is a promising computing paradigm that can alleviate this bottleneck by reducing processor--memory data transfers and enabling massively parallel computation within DRAM (e.g., across 65,536 columns). However, weight matrices in individual LLM layers (e.g., a QKV matrix in Llama3 8B with 6,144 rows) are typically too small to fully utilize parallelism in PUD. To address this challenge, we propose an algorithm--system co-design for PUD-based LLM inference. At the algorithm level, we enable a subsequent projection to be packed with the current projection by reformulating it as a computation on the current residual state plus a correction for the residual update. Layer redundancy allows this correction to be approximated using fewer bit-serial steps. At the system level, we tailor data mapping and execution to improve utilization of either columns within DRAM subarrays (SIMD) or independently controlled banks (MIMD). Overall, we achieves up to $1.41\times$ end-to-end decoding speedup over the PUD baseline, corresponding to a $2.45\times$ speedup over processor-based execution.

Apto: Self-Adaptive CPU-GPU Hybrid MoE Inference on Shared PCs

Zihang Zhong, Jiali Wang, Rong Chen — Tue, 03 Nov 2026 00:00:00 +0000

94% · Inference speed 94%, Serving systems 64%

CPU/GPU hybrid inference enables Mixture-of-Experts (MoE) models to run on memory-constrained PCs by keeping all expert weights in host memory and executing activated experts across both processors. Existing hybrid systems, however, assume exclusive resource access and determine an upfront CPU/GPU partition, so CPU contention from colocated applications can severely imbalance the two execution paths. This paper presents Apto, a self-adaptive hybrid inference system that applies dynamic scheduling to expert execution. The CPU and GPU executors claim experts from opposite ends of a dual-ended queue as they progress, allowing the CPU/GPU work split to adapt to their execution progress without predicting contention. Apto further uses affinity-aware scheduling to expose each executor's preferred experts at its queue end and cancellable tail racing to mitigate final-expert stragglers. Compared with state-of-the-art GPU-only and hybrid baselines, Apto achieves average prefill and decode speedups of 2.96$\times$ and 1.52$\times$ under standalone execution, and 5.97$\times$ and 3.45$\times$ under concurrent CPU workloads. It also accelerates real-world coding-agent applications by 2.02$\times$ on average.

Quantizing Recurrent Language Models Is Easier Than You Think

Philipp Nazari, Neehal Tumma, Daniela Rus, T. Konstantin Rusch — Tue, 03 Nov 2026 00:00:00 +0000

94% · Compression 94%, Inference speed 79%

Linear recurrences are becoming an increasingly common building block in large language models, replacing the growing KV cache of softmax attention with a fixed-size recurrent state that enables efficient decoding. However, the quantization stack has largely been developed around softmax Transformers, whose numerical properties and inference state differ from those of recurrent models. Our goal is to understand these differences and study the design space of quantization methods for recurrent models. In our analysis, we differentiate between two targets. First, extra-state quantization concerns weights and activations outside the recurrent state, whose distributions are shaped indirectly by the sequence mixer during training. In controlled 1.3B parameter models, we find that decayed recurrences develop weaker residual outlier channels than softmax attention and are therefore easier to quantize without incoherence processing. Second, intra-state quantization concerns the recurrent state itself during decode, where repeated updates create a different numerical problem from KV cache quantization. We instantiate our findings as LoReS, a state quantization pipeline combining value-axis scale grouping to match output channel magnitude, decay-based mixed-precision to protect slowly decaying heads or channels, and stochastic rounding to limit error accumulation under repeated requantization. Across Qwen-3.5 9B, Nemotron-H 8B, and Kimi-Linear 48B-A3B, LoReS supports robust 8-bit state quantization and lower effective precision on shorter decode horizons. We implement LoReS in vLLM and demonstrate end-to-end decode speedups.

PagedMoE: Cross-Precision Expert Caching for Token-Level Dynamic Precision in SSD-Resident MoE Inference

Tue, 03 Nov 2026 00:00:00 +0000

94% · Inference speed 94%, Serving systems 73%, Compression 37%

Large Mixture-of-Experts (MoE) models are promising for local inference because sparse routing allows the routed-expert pool to reside on SSD while DRAM caches recently used experts. However, under practical DRAM cache capacities, substantial SSD I/O remains on the critical path despite I/O-computation overlap. Token-level dynamic expert precision offers a natural way to reduce this data demand by allocating bits according to expert importance and quantization sensitivity. Nevertheless, treating each bit-width of the same expert as an independent object prevents cross-precision cache reuse, causing overlapping data to be fetched repeatedly. The key insight is that precision should instead denote the refinement depth of one shared expert state. Based on this insight, we present PagedMoE, which stores each expert once using a prefix-compatible tree codebook and bitplane layout. Lower-precision representations are valid prefixes of higher-precision ones, allowing cached data to be reused or extended by fetching only missing bitplanes. PagedMoE selects the required prefix length before I/O using current routing weights and offline expert-specific error profiles, and directly executes cached 1-8-bit prefixes without materializing full-precision experts. At an average routed-expert precision of 2.5 bits, PagedMoE achieves quality close to GPTQ-INT4 on Qwen3.5-397B-A17B and to MXFP4 on most evaluated GLM5.2 tasks. With a 24 GiB expert cache, it reaches 8.89 tokens/s on Qwen3.5, 1.33x the throughput of the same dynamic requests using independent precision objects and 1.56x that of uniform 4-bit execution.

Output-Aware Block Influence: Jacobian-Lens Weighting for Depth Pruning

Tue, 03 Nov 2026 00:00:00 +0000

93% · Compression 93%, Inference speed 78%

Depth pruning removes entire transformer blocks and therefore depends on a reliable criterion for identifying which blocks can be removed. Block Influence (BI) ranks blocks by how minimally they rotate the hidden state. However, local redundancy does not necessarily imply downstream irrelevance. We address this limitation using the Jacobian Lens (J-Lens) neural artifact, a map transporting perturbations at a layer to the final residual stream. At a common sparsity level, pruning scores applying J-Lens transportation to $BI$, what we call $J-BI$, were found to improve post-pruning accuracy by up to 19 points over $BI$ while reducing damage to perplexity for Qwen3-8B, Llama-3.1-8B-IT, and Gemma-3-12B-IT. We also perform a single-block ablation study and use J-Lens probing to explain why the proposed scores improved performance.

Quantization of Large Models with Weight Outliers and Sparsity

Tue, 03 Nov 2026 00:00:00 +0000

93% · Compression 93%, Inference speed 77%

Post-training quantization (PTQ) reduces the memory burden of large model deployment by casting the high precision trained weights to fewer bits. While this promotes memory-efficient inference, it also incurs quantization error that impairs model performance. To compensate for quantization error, the present contribution decomposes the original trained weights into a quantized matrix, a low-rank component, and a sparse matrix that accounts for outlier rows in the weights. Due to analytical intractability and the discrete structure of the decomposition, a simulated annealing search is employed to explore candidate decompositions while avoiding poor local minima. Numerical tests demonstrate the effectiveness of this approach across a broad suite of large language models, including Llama 3.2 1B, Llama 3.2 3B, Llama 3.1 8B, and Qwen2.5 7B, evaluated on natural language generation benchmarks using 4-bit NF4 and FP4 quantizers. The novel method can markedly outperform direct quantization while incurring minimal memory overhead.

CacheWeave: Learning to Recompute for Efficient Position-Independent KV Caching

Yueyang Pan, Qianchao Zhu, Yuliang Liu, Teng Ma, Sanidhya Kashyap — Tue, 03 Nov 2026 00:00:00 +0000

93% · Inference speed 93%, Serving systems 37%

Position-independent caching reuses chunk KV caches across positions and requests. Independent encoding, however, omits cross-chunk interactions and can degrade accuracy. In our study, neither PIC training nor rule-based recomputation, alone or combined, consistently recovers full-prefill accuracy. Oracle-guided recomputation on PIC-trained models improves accuracy at the same budget, motivating better selection. We present CacheWeave, which adapts the full backbone to PIC, then freezes it and trains a model-native block selector on the adapted model's oracles. The selector uses compact cached chunk features and question hidden states without extracting backbone attention scores. It selects contiguous token blocks aligned with KV blocks to support efficient recomputation and limit copy-on-write amplification. Its runtime, CacheWeave-Serve, supports private and shared KV caches, keeps selection and recomputation on the GPU, and stages sliding-window KV for hybrid attention models. Across three models and four tasks, CacheWeave outperforms CacheBlend and ProphetKV by 7.2 and 3.5 percentage points on average at a 5% recomputation budget, with a mean relative accuracy loss of 1.6% to full prefill. With 32-token blocks, CacheWeave-Serve cuts single-request TTFT by up to 16.1x over full prefill and 1.7x over ProphetKV, and raises SLO goodput under load by up to 7.8x.

ReFrame: Silence-Aware Compute Reclamation for Full-Duplex Speech Models

Zheng Zhou, Guanyi Chen, Jing Wei, Haoyuan Liang, Jian Zhang, Haohuan Fu — Tue, 03 Nov 2026 00:00:00 +0000

93% · Inference speed 93%, Serving systems 49%

Full-duplex speech models execute dense model inference for every audio frame throughout a live interaction. This uniform framewise schedule creates sustained GPU load across dialogue states whose computation has strong, reusable temporal structure. ReFrame formulates full-duplex inference as a silence-aware computation problem. Mutual Silence Speculation (MSS) operates when both participants are silent: it evaluates multi-frame Temporal continuations, verifies text and hidden-state stability, and applies batched Mimi audio-token correction. Listening Frame Folding (LFF) operates while the user speaks and the agent listens: it processes already-arrived user frames through parallel Temporal execution and frame-batched audio-token generation while accepting only the verified prefix. Frame-Budget Reinvestment (FBR) assigns the released frame intervals to resumable external retrieval and language-model workers, then injects timely references through Moshi's conditioning interface. We instantiate ReFrame on five full-duplex models spanning Moshi-family, phase-structured, and flattened autoregressive generation. Across the complete 1,225-conversation FDBench, ReFrame reduces measured GPU execution time by 18.80-31.01% and retains 96.92-100% non-PAD text agreement on five different models. On HaluEval tests, FBR increases the number of correct generated spoken answers from 86 to 205 under the same auxiliary-service budget. ReFrame turns recurring dialogue structure into lower inference cost and higher end-to-end service capability.

Introducing Cluster Efficiency to LLM Serving: Pooling Economical Accelerators over Fast Interconnects

Youhe Jiang, Pavle Glu\v{s}ac, Binhang Yuan, Eiko Yoneki — Tue, 03 Nov 2026 00:00:00 +0000

93% · Inference speed 93%, Serving systems 77%

Long-context large language model (LLM) serving places substantial demands on compute, memory capacity, and memory bandwidth. Economical accelerators can offer these resources at lower cost, but their limited per-device capacity requires pooling across devices, making communication overhead critical. We study cluster efficiency, measured in tokens per dollar and per joule, and show how low-latency, high-bandwidth interconnects enable economical accelerators to retain their cost advantage. Guided by a characterization of coding-agent workloads and accelerator scaling, we realize this design on the Tenstorrent Galaxy, a cluster of Blackhole chips. The serving stack combines integrated communication engines, pipelined ring transfers, communication-computation overlap, and device-side decode replay to reduce collective overhead. Across three dense models in BF16 and 8-bit configurations with 8K-256K-token contexts, the Galaxy achieves 1.07-2.24$\times$ the tokens per dollar of an eight-GPU RTX PRO 6000 Blackwell server under a three-year acquisition-and-electricity cost model. Emulating TCP-class communication on the same chips reduces throughput by 23.5-67.4% and eliminates the cost advantage in every evaluated configuration. These results demonstrate how fast interconnects translate the per-dollar resource advantages of economical accelerators into cost-efficient long-context LLM serving.

TVP: TV is Almost All You Need for Speculative Decoding

Aaryam Sharma — Tue, 03 Nov 2026 00:00:00 +0000

92% · Inference speed 92%

Speculative decoding accelerates LLM inference with a drafter whose proposed tokens are verified by the target model in parallel. Under lossless speculative sampling, a draft token is accepted with probability one minus the total variation (TV) distance between the target and draft distributions, yet drafters are often trained with cross-entropy (CE) or KL. We first show that under matched and well-tuned training, TV matches or surpasses several published draft-training objectives when fine-tuning, and a fixed TV and KL mixture does the same when training from scratch. We identify two remaining problems, TV's vanishing gradients on undercovered tokens and the prefix products that starve later positions in acceptance-length objectives. To address both, we derive TVP, a loss family based on expected acceptance length, whose power parameter interpolates between coverage and acceptance oriented objectives. Annealing the power from zero to one gives the highest or tied-highest acceptance length across three drafter architectures and three target models, including longer drafts and block verification. TVP improves acceptance length over CE by 9.8% when fine-tuning, and over the public DFlash2 training recipes by 6.2 to 12.9%.

Scute: Exact Batched Serving for PEFT Adapters Beyond LoRA

Thomas Walker, Richard Baraniuk — Tue, 03 Nov 2026 00:00:00 +0000

92% · Serving systems 92%, Inference speed 80%

Existing serving systems for large language models batch requests for many different LoRA adapters on one GPU. However, these systems reject adapters from other parameter-efficient fine-tuning (PEFT) methods such as DoRA, (IA), VeRA and OFT. One solution to this is to convert these adapters to LoRA, but this induces a significant memory and accuracy cost. Instead, we show that, under some modifications, a batched LoRA kernel can serve all of these methods exactly and efficiently. We achieve this by writing each method in one normal form, a LoRA computation with six slots for an adapter's own vectors and matrices. Some slots can merge into an adapter's weights at load time; the others remain live and induce a cost. We give ways to efficiently pay this cost and predict it from a zero-parameter byte model. We package this system, Scute, and build it on top of vLLM. Scute serves all five methods, coming as checkpoints from the Hugging Face PEFT library, in one batch, and it matches that library with a mean error of at most 1.5% of each adapter's effect. Under a burst of requests on an RTX A6000, a mix of these methods takes 5-17% more decode time than the same number of LoRA adapters. Furthermore, using Scute, the same GPU can serve 64 to 1024 (IA) and VeRA adapters at 1.5x to at least 2.5x the request rate of their LoRA conversions.

FactorSpec: Serving Exact Relational Constraints as Speculative Drafts

Jianchang Su, Dawei Xiang, Wei Zhang — Tue, 03 Nov 2026 00:00:00 +0000

92% · Inference speed 92%, Serving systems 67%

Speculative decoding speeds up LLM serving by drafting tokens that the target model verifies in one pass, and its gain grows with the drafts the target accepts. Structured outputs should be an ideal case, because much of a record follows from relations among its fields: an itinerary's days follow from its cities' stays and flights. Serving engines, however, draft such text by proposing it from the prompt or a learned head, and their grammar masks enforce only syntax, so the model must satisfy the relations, which it violates: on NATURAL PLAN's trip problems, Qwen3-8B writes a valid itinerary for 10%, and retries that name the violated relation raise this to 14%. We present FactorSpec, which serves exact relational constraints as speculative drafts. FactorSpec compiles a JSON Schema and its relations into a factor graph with an exact feasibility check, drafts every token the constraint determines as certain, verifies each request at a width chosen for it, and computes the masks in worker processes outside the scheduler while the GPU runs. In SGLang, its drafts yield up to 10.5 tokens per verify step, up to 1.9x those of the fastest n-gram drafter, and it decodes up to 1.43x faster than speculators tuned per load, at a grammar mask's cost; it also leads in vLLM. Because every record is also valid, it serves up to 13.9x the valid records per GPU-second of syntax-only baselines.

SHq: Learnable Grids for LLM Weight Quantization

Michael Helcig, Earl Killian, Dan Alistarh — Tue, 03 Nov 2026 00:00:00 +0000

92% · Compression 92%, Inference speed 45%

Block-scaled lookup-table (LUT) formats, which represent weights as indices into a small value set (the grid), have become standard for low-bit LLM quantization, ranging from NF4 and BOF4 to the 3-bit LUT instruction in NVIDIA Rubin. While these formats leave the grid definition flexible, existing implementations rely on fixed grids derived either from numerical data types (INT, FP) or assumed weight distributions (NF, BOF). We demonstrate that no single fixed grid is universally optimal; the best fixed grid varies with bit-width: BOFS at 2 bits, NF at 4 and 5 bits, and INT at 6 bits. We introduce SHq, a closed-form family of grids parameterized by two critical degrees of freedom: a companding strength {\alpha} that interpolates level spacing from uniform to floating-point-like, and an offset c that introduces asymmetry. This offset allows the sign bit of the FP8 block scale, otherwise unused, to select between the grid and its mirror image, up to doubling the number of distinct values available per block. SHq modifies only the lookup table contents and serves as a drop-in replacement for the grid in any such format. Across five models ranging from 3B to 35B parameters, including a mixture-of-experts architecture, evaluated at 2-6 bits using RTN and GPTQ, SHq achieves the lowest KL divergence in every setting at 3 and 4 bits and in most settings at 2, 5 and 6 bits. It outperforms the best fixed grid by up to 34% and remains superior under mixed-precision configurations. At 4.5 bits per weight, the SHq format reduces the KL divergence of NVFP4 by 41-67% while improving downstream accuracy. Grid families incorporating both spacing and offset parameters perform comparably, whereas those lacking an offset exhibit clearly higher error, confirming that these two degrees of freedom drive the observed gains. We therefore propose SHq as the standard grid for block-scaled lookup-table formats.

C3: Speech-Guided KV Retrieval for Long-Context SpeechLLM Inference

Tue, 03 Nov 2026 00:00:00 +0000

91% · Inference speed 91%

Speech language models encode audio at a much higher rate than text, causing the KV cache to dominate the memory, bandwidth, and decoding cost of long-form generation. KV retrieval can reduce per-step access by attending to only a query-dependent subset of the cache, but existing methods are designed primarily for text and do not account for the structure of audio attention. We first show that retrieval errors may compound autoregressively into generation drift. We then analyze full audio attention during long-form generation and identify three properties relevant to retrieval: unequal attention concentration across query heads, non-contiguous audio support, and temporal continuity across decoding steps. Guided by these observations, we propose $C^3$(Cluster-Compete-Compete), which organizes audio KVs in a hierarchical key-space index, lets query heads sharing a KV head compete for a common retrieval budget, and adaptively reallocates retrieval resolution through paired split-merge updates. Across long-form speech recognition, translation, and summarization on four SpeechLLMs, $C^3$ achieves the strongest overall performance among retrieval methods under comparable final-KV access budgets, with particularly large gains in the tight-budget regime, while improving long-context decoding efficiency and accelerator-memory scaling.

RankLane: Efficient MoE Prefill on Accelerators with Ranked Expert Capacities

Wentao Ni, Chihao Yu, Hong Cai, Mingu Lee, Hejia Zhang, Fatih Porikli — Tue, 03 Nov 2026 00:00:00 +0000

91% · Inference speed 91%, Serving systems 83%

Mixture-of-experts (MoE) models expand model capacity by activating only a subset of experts for each token. On accelerators with static execution plans, expert capacities and placement are fixed before routing reveals each expert's load. Uniform worst-case capacities waste work on padding; smaller capacities tied to expert identities can overflow. We present RankLane, which ranks each device's experts by current load and assigns them to fixed-capacity lanes. Runtime indices select experts and token rows while tier populations, shapes and device placement remain fixed. An aggregation graph preserves device ownership and reduces each token's local contributions on one core. Routing calibration and measured execution costs guide tier selection; overflowing chunks rerun at full capacity. We evaluate full-model prefill of Qwen3-30B-A3B and Gemma 4 26B-A4B with MXFP6 weights on four Cloud AI 100 Ultra SoCs, using 128-2048-token chunks and including overflow detection and in-program recovery support. Against a dense baseline, RankLane achieves end-to-end prefill speedups of 1.27-4.18x on Qwen3 and 1.93-3.46x on Gemma.

LoRAForge: Co-Designing Post-Hoc Compression and Execution for Efficient Multi-LoRA Serving

Shenghui Li, Thiemo Voigt — Tue, 03 Nov 2026 00:00:00 +0000

91% · Inference speed 91%, Serving systems 84%

Multi-LoRA serving allows many task-specific adapters to share a common base model, but scaling to large adapter fleets introduces substantial storage, loading, and execution overheads. Post-hoc compression can reduce the memory footprint of trained LoRA adapters, yet smaller ranks do not automatically translate into faster serving. For instance, padding, factor access, and generic kernels can obscure the retained low-rank structure, while independently recompressing thousands of modules also incurs significant preparation overhead. We present LoRAForge, a system connecting batched factor construction to rank-specialized execution through an explicit retained-rank representation. It supports two allocation policies: uniform retention ratios (LoRAForge-U) and spectral-energy budget allocation (LoRAForge-S). Both share FlashTSQR factor construction and the same serving backend. FlashTSQR exploits parallelism across modules and row tiles, accelerating factor construction by up to $2.28\times$ over matched library QR. In a fixed-work Qwen control, the shared runtime adds $1.17$--$1.22\times$ throughput across both compressed policies, with negligible gains on original adapters. Across 85 trained open-weighted Qwen and Llama adapters, the spectral policy's floored quality configuration retains about 7.7% of adapter parameters with fleet-average score drops of at most 1.2 points on a 0--100 scale. The same rank-floored $1/12$ spectral exports achieve substantial token-throughput gains on both AMD and NVIDIA GPUs.

Integer-Exact Microscaling: Running MXFP4 Language Models on INT8 Matrix Units

Shinyoung Kang, Junhyung Ahn, Yujun Kim, Jungho Park — Tue, 03 Nov 2026 00:00:00 +0000

90% · Inference speed 90%, Compression 85%, Serving systems 64%

On hardware whose fastest matrix datapath is INT8$\times$INT8, such as the Neural Accelerators of Apple's M5 GPUs, 4-bit weights must become INT8 operands. Existing MXFP4 kernels already map FP4 (E2M1) values to integers by doubling them, but apply every 32-element block scale inside the reduction. We show when no per-block work is needed: a row of MXFP4 weights is exactly one INT8 vector times one power-of-two scale if its block exponents span at most three, so a linear layer becomes a single INT8 GEMM with one rescale per output. Enforcing the condition adds at most 1/16 of a row's largest rounding error to the blocks it changes, and real weights already satisfy it: 97.9-99.97% of rows across seven models, including 99.93% of the 95M rows of a 125B-parameter mixture-of-experts (MoE) model, Qwen3.8-Flash-Next. The resulting spread-constrained MXFP4 files stay standard MXFP4. With integer instead of power-of-two block multipliers, an integer-exact format (16-wide blocks, multipliers 1..10, 4.26 bits) runs as one exact INT8 GEMM and is more accurate than NVFP4 (4.5 bits) with its standard scale rule on five small models, within 0.06 perplexity of NVFP4 with scale search. On Qwen3.8-Flash-Next's routed experts with INT8 activations it is as close to the FP8 model as NVFP4 with scale search (KL divergence 0.0557 vs. 0.0553) and closer than MLX's MXFP4 and affine INT4; with every matrix quantized it solves 94.8% of GSM8K vs. 91.0% for MLX's MXFP4. Quantizing only the 74k routed expert matrices (96% of the parameters) to either of our formats with INT8 activations costs at most 0.2% perplexity. A complete W4A8 system on the M5 produces outputs bit-identical to the quantized reference; prefill is 1.34-1.74x faster than MLX on a dense model, 1.10-1.57x on two MoE models up to 35B parameters and 1.22-1.29x on Qwen3.8-Flash-Next, where shared non-GEMM work bounds the gain.

Fairway: Load-Aware KV Cache Reuse for LLM Prefill Scheduling

Tue, 03 Nov 2026 00:00:00 +0000

90% · Inference speed 90%, Serving systems 53%

LLM-based applications often employ agentic workflows and retrieval-augmented generation (RAG), which build long prompts from shared context. When prefixes are shared across requests, serving engines reuse the key-value (KV) cache to skip parts of the prefill computation. Reuse is typically graded: the uncached work of requests in a queue can differ by up to 100x. In this setting, the order of scheduling determines the time-to-first-token (TTFT) for the queued requests. KV reuse-first scheduling, which prioritizes the requests with the longest KV cache hit, gives the lowest median TTFT but starves expensive requests under load spikes; first-come-first-served (FCFS) scheduling bounds the tail but raises the median TTFT. We present Fairway, a prefill scheduler built on adaptive reuse prioritization: it favors KV cache reuse while queues are short, and shifts toward fair allocation as requests wait longer relative to their own prefill time. Fairway employs a virtual dual lane model: a KV reuse-first Fast lane and a Fair lane atop the serving engine's waiting queue that enables speeding up of cheap reuse requests without starving expensive ones. This model automatically adapts to the workload without any per-workload tuning. Fairway lowers the tail TTFT at no cost to the median in transient overload. Across workloads and load levels, Fairway matches or beats the best baseline, with a p99 TTFT up to 8.6x lower than prior prefix-aware schedulers. During load spikes, it keeps the tail of FCFS at a 6x lower median TTFT.

Concise-9B: Removing the Slack from a Reasoning Model's Thinking

Alexander Boesgaard Lorup — Tue, 03 Nov 2026 00:00:00 +0000

90% · Inference speed 90%

Reasoning models answer better when they think longer, but a thinking trace keeps going well after its answer has formed. We measure this slack in Qwen3.5-9B and remove most of it with concise self-training: two rounds of on-policy rejection-sampling fine-tuning that keep the shortest correct samples, plus a small share of answer-synthesis targets, giving Concise-9B. On 8,983 held-out problems (competition and hard mathematics, SuperGPQA, a five-area confirmation set and MMLU-Pro), Concise-9B generates 0.66x the tokens of Qwen3.5-9B at a 32K thinking budget, with accuracy equivalent within +/-1 point (-0.20 [-0.77, +0.36] points). Forcing each trace on the hard and confirmation sets to answer every 2,048 thinking tokens, we find that 54-64% of Qwen3.5-9B's thinking tokens come after the point from which every forced answer equals the final one (41-48% for Concise-9B). Stopping there would keep every answer at 0.35-0.44x the base model's tokens, but finding that point takes hindsight: stopping when three consecutive forced answers agree costs the base model 2.4-9.2 points. Concise self-training removes slack without that cost: 86% [83, 90] of the thinking tokens it saves come after the answer has formed. It has a limit: with traces continued to 80K tokens, Concise-9B solves 25 fewer of 1,276 hard problems (-25 [-49, -1]), mostly in competition mathematics. The measurement also yields a cheap commit signal for adaptive sampling that matched two-sample escalation within +/-1 point at 0.67x its tokens in a pre-registered test. Results on Qwen3.5-4B, on which the recipe was developed, and a replication on Olmo-3-7B-Think suggest that the saving depends on how verbose the base model is.

★ HiState: Hierarchical Prefix Cache Management for Hybrid LLM Serving in Agentic Workloads

Zaifeng Pan, Lucas Wang, Xinwei Qiang, Yanbo Zhou, Chang Chen, Zhengding Hu — Mon, 02 Nov 2026 00:00:00 +0000

96% · Inference speed 96%, Serving systems 88%

Hybrid large language models combine full-attention and recurrent linear layers to reduce inference costs. Prefix caching can further reduce redundant computation across requests, but a recurrent checkpoint cannot recover an earlier state corresponding to a shorter prefix. Existing hybrid prefix caching systems retain checkpoints at selected boundaries to support common reuse patterns in conventional conversations. However, agent workloads modify and revisit earlier contexts, exposing reusable prefixes whose checkpoints were never saved or have already been evicted. A straightforward approach is to retain more checkpoints to reduce recomputation, but their cumulative footprint can displace reusable prefixes from other sessions and cause additional cache misses. To address these problems, we present HiState, a prefix cache management system for hybrid models serving agent workloads. Our key observation is that prefix reuse accesses all matching key-value (KV) entries but only one recurrent checkpoint, regardless of how many checkpoints are retained. This asymmetry allows HiState to retain historical checkpoints on SSD while prioritizing fast memory for KV entries: each hit retrieves only one checkpoint whose size is independent of context length. Within this hierarchy, we propose adaptive interval control to balance checkpoint storage against recomputation, and unified eviction to coordinate checkpoint sparsification with reclamation of stale trajectories. We implement HiState on top of vLLM and evaluate it on two public coding-agent trace datasets and controlled microbenchmarks. On real agent traces, HiState achieves up to 5.73x and 5.44x speedups in p99 TTFT over vLLM-Marconi and SGLang-HiCache, respectively.

★ CAMEL: Cross-Architecture Masked Embedding Learning for Ultra-Lightweight Retinal Foundation Models

Mon, 02 Nov 2026 00:00:00 +0000

95% · Compression 95%

Recent progress in deep learning has substantially advanced automated grading of ophthalmologic pathologies, supported by large annotated datasets and improved representation learning. However, a significant gap remains between state-of-the-art performance and clinical scalability. In particular, foundation models have reshaped image-based medical screening and diagnosis. These large-scale architectures trained on extensive datasets learn rich, generalizable feature representations that can be adapted to a wide range of downstream tasks. In ophthalmology, RETFound, one of the most relevant foundation models for retinal images, performs strongly across classification, segmentation, and risk prediction tasks. Despite their effectiveness, foundation models are often computationally prohibitive, hindering their deployment in real-world clinical environments where hardware constraints and latency are critical. To bridge this gap, we introduce CAMEL (Cross-Architecture Masked Embedding Learning), a novel cross-architecture distillation framework that uses an asymmetric masking strategy during embedding learning to effectively transfer knowledge from massive teacher foundation models to compact, lightweight student models. Our experimental evaluation across four datasets shows that CAMEL preserves accuracy while drastically reducing computational complexity, enabling lightweight yet highly accurate models suitable for resource-limited clinical settings.

★ Trainable Vector Quantization for Large Language Models

Mon, 02 Nov 2026 00:00:00 +0000

95% · Compression 95%, Inference speed 78%

Vector Quantization (VQ) is a rapidly evolving technique for compressing Large Language Model (LLM) weights, reducing memory usage, and speeding up inference. However, extreme compression deteriorates model performance, which is challenging to recover due to the discrete nature of VQ parameters. Existing methods employ training-free approaches that are limited to layer-wise reconstruction loss minimization. To address this limitation, we propose TVQ, a novel method that leverages straight-through estimation to enable gradient descent-based optimization of vector-quantized LLMs. By reformulating VQ LLM optimization with trainable latent weights, TVQ enables arbitrary objective minimization. We constitute our approach by applying TVQ in a block-wise knowledge distillation paradigm, which balances quality and training efficiency, and propose several practical solutions to accelerate training. In experiments on modern LLMs, we demonstrate that TVQ is compatible with modern VQ schemes. As a result, TVQ achieves state-of-the-art quality in low-bit post-training quantization settings with a modest training budget and straightforward PyTorch implementation. For instance, TVQ achieves near-lossless 2-bit quantization for LLaMA2-70B with only 1% accuracy drop on popular Question-Answering benchmarks.

SparseGDN: Chunk Sparse in Linear Attention

Jiedong Zhuang, Peng Zhao, Jialin Tang, Haoji Hu — Mon, 02 Nov 2026 00:00:00 +0000

93% · Inference speed 93%, Serving systems 69%

Gated Delta Networks (GDNs) offer linear-time sequence modeling, yet long-context prefill remains costly due to dense intra-chunk computation, recurrent-state updates, and intermediate state storage. We introduce SparseGDN, a hardware-aware framework that combines structured intra-chunk approximation with input-dependent recurrent-event selection while producing outputs for every token. A block-adjacent solver reduces local Delta-solve computation, while a dynamic selector allocates recurrent-update budgets across head pairs. The GPU implementation materializes updates only for selected events and reconstructs token outputs using compact state snapshots and group-local interactions, reducing computation and memory traffic. We evaluate SparseGDN on Qwen3.5-4B and Qwen3.5-9B across RULER tasks and context lengths from 4K to 128K. Retaining approximately 29% of recurrent events on average, SparseGDN achieves 1.50-1.94x GDN-core prefill speedups over native FLA kernels, including online selection overhead.

★ MASS: Multi-mask Selective Self-Distillation for Leakage-Resilient Reasoning

Xiuqi Liu, Oveys Delafrooz Noroozi, Upamanyu Madhow — Fri, 30 Oct 2026 00:00:00 +0000

96% · Compression 96%

On-policy self-distillation enables the same model to act as student and teacher, with dense token-level supervision obtained by the teacher being given access to privileged information. However, conditioning the teacher on a single reference can concentrate its predictions on one reasoning trajectory and limit transfer to the student. We propose Multi-Mask Selective Self-Distillation (MASS), which approximates the unavailable marginal over diverse reference realizations using multiple independently masked views of one verified solution. MASS aggregates these complementary teacher predictions, selectively transfers reliable token-level corrections, and adaptively focuses masking on influential reference spans. Across mathematical reasoning benchmarks and model families, MASS consistently outperforms strong self-distillation baselines.

★ On-Policy Self-Distillation with Teacher Evolution

Shufan Shen, Zhongni Hou, Chao Bi, Yufei Zhang, Guojun Yin, Shuhui Wang — Fri, 30 Oct 2026 00:00:00 +0000

96% · Compression 96%

On-Policy Self-Distillation (OPSD), which aligns a student model's token-level distribution with that of its self-teacher prompted by privileged context, has emerged as a promising direction to mitigate reward sparsity in sequence-level post-training paradigms. However, OPSD often fails to achieve continuous performance gains during training. In this paper, we find that the model's limited self-teaching capability underlies this phenomenon by generating unreliable guidance on certain samples. To address this problem, we propose Teacher-Evolution Self-Distillation (TESD), which enables the self-teacher to generate rollouts and receive environmental feedback based on task correctness and context independence. First, using these rewards as training objectives, we perform teacher-oriented optimization to continuously enhance the model's self-teaching capability during distillation. Second, by estimating teacher guidance reliability from these rewards, we apply adaptive student-oriented optimization to down-weight samples with unreliable guidance. Extensive experiments across multiple datasets and model scales show that TESD consistently outperforms representative token-level and sequence-level methods, highlighting the necessity of improving supervision quality in OPSD.

★ PDScope: Breaking the Prefill-Decode Seesaw for Memory-Bound MoE Inference

Enda Yu, Dezun Dong, Zhaoning Zhang, Xiaojian Yang, Weiling Yang, Shaonan Ma — Fri, 30 Oct 2026 00:00:00 +0000

96% · Inference speed 96%, Serving systems 85%

Local deployment of MoE-based LLMs on CPU-GPU nodes is attractive for cost and privacy reasons, yet limited GPU memory forces most experts into CPU memory. Prefill is then compute-bound and benefits from aggressive GPU loading, while decode is I/O-bound and requires conservative PCIe transfers. These demands cannot be amortized under low batching---typical of local assistants and document tools---forcing phase-agnostic schedulers into a seesaw: favoring prefill's aggressive loading injects unmaskable I/O that stalls decode, while favoring decode's conservative transfers starves prefill's GPU of work. We observe that expert routing is predictable in phase-specific ways, allowing us to decouple the two phases and optimize prediction, scheduling, and execution independently for each phase. PDScope predicts expert activations to guide placement and prefetching, adapts these policies to each phase's bottleneck via cross-layer cost modeling, and compiles decisions into overhead-free pipelines using a static computation graph, cache prefetching, and zero-copy asynchronous I/O. On mainstream MoE models, PDScope achieves 2.5-10.3x higher prefill throughput and 2.1-6.1x higher decode throughput than existing systems, showing that phase-aware co-design unlocks efficient MoE inference under memory constraints and extends naturally to multi-batch settings.

★ FlatCake: A Flat KVCache Storage Architecture for LLM Serving

Xin Huang, Teng Ma, Zhuoran Li, Yibo Zhao, Zhaoshi Li, Qianchao Zhu — Fri, 30 Oct 2026 00:00:00 +0000

95% · Inference speed 95%, Serving systems 82%

KV cache reuse accelerates LLM serving, but existing systems manage host DRAM and NVMe SSD as a tiered hierarchy where DRAM acts as an intermediate cache. Production LLM workloads break both premises of this hierarchy. First, bimodal reuse defeats DRAM temporal filtering, since workloads are dominated either by one-shot blocks with long reuse distances or by multi-gigabyte agentic sessions that exceed DRAM capacity, putting SSD on the critical read path. Second, DRAM staging and eviction-ordered layout bottleneck SSD restoration. Holding host staging buffers for in-flight prefixes causes contention under load that aborts prefetches, while incremental eviction scatters a session's blocks across the SSD and prevents large sequential reads. We present FlatCake, a flat KV storage architecture that exposes host DRAM and NVMe SSD as peer media in a unified 64-bit address space. Exploiting the fact that KV blocks are write-once, immutable, and reused at session scope, FlatCake places each block directly onto its target medium in session-contiguous, layer-major order at creation time and never migrates it. FlatCake streams coalesced SSD reads directly into a bounded GPU window via GPUDirect Storage and decouples DRAM and SSD request queues to prevent head-of-line blocking. On a production Claude Code trace with DeepSeek-V4-Flash, FlatCake sustains a 94% cache hit rate as load increases (versus 64% for Mooncake), eliminates Mooncake's $6.5\times$--$6.7\times$ SSD read amplification, and reduces peak-load mean TTFT by $1.6\times$ while improving total token throughput by $1.6\times$. Across 8K--128K contexts on Llama-3.1-70B, FlatCake further reduces mean TTFT by up to $2.4\times$ over Mooncake and $1.8\times$ over Mooncake-GDS.

★ A Unified Hessian-Guided Framework for Post-Training Quantization

Mads Toftrup, Marco Federici, Boris van Breugel, Paul Whatmough, Markus Nagel — Fri, 30 Oct 2026 00:00:00 +0000

95% · Compression 95%, Inference speed 69%

Low-bit quantization enables efficient deployment of large language models by reducing memory footprint and computational cost. Achieving high accuracy at low precision requires selecting quantization ranges, linear transformations, and weight rounding decisions. Existing approaches typically optimize these components independently using local reconstruction objectives, which fail to account for the varying impact of quantization errors on downstream predictions. Direct optimization of task-level metrics would provide a more faithful objective, but is challenging due to the discontinuous nature of quantization operations and the high computational cost of evaluating their downstream effects. In this work, we derive a Hessian-guided approximation of the output change induced by joint weight and activation quantization. Based on a tractable Fisher approximation, we formulate a unified post-training quantization framework that enables the joint optimization of quantization ranges, linear transformations, and weight rounding under a common output-sensitivity objective. The resulting method remains practical for modern large language models and consistently outperforms state-of-the-art transformation and rounding methods across both channel-wise and block-wise quantization settings.

Causal Attention on CPU: A Cache-Aware and Load-Balanced Scheduling Kernel for Prefill

Kwun Fung Lau, Frost Mitchell, Kanya Mo, Emani Ashok — Fri, 30 Oct 2026 00:00:00 +0000

94% · Inference speed 94%, Serving systems 89%

Agentic AI drives significant productivity gains across diverse domains, but concurrent long-context prefills from multiple agents can saturate GPU compute resources. Although CPUs offer a viable alternative, existing CPU attention kernels fall short in adapting to multi-tier cache hierarchies and lack effective thread scheduling. To address this, we introduce CACPU, a high-performance attention kernel featuring tiered cache blocking and an in-kernel task dispatching mechanism designed for causal full-sequence prefilling. CACPU achieves up to a 2.01x speedup over PyTorch SDPA, FlexAttention, and SGLang attention kernels. Integrated into SGLang, it reduces Time to First Token (TTFT) of Llama-3.1-8B by up to 1.74x while preserving model accuracy.

DWISE: Denoising Workload-Informed Scheduling Engine

Lie Li, Wen Li, Junxiao Shen, Jindong Gu, Guosheng Hu — Fri, 30 Oct 2026 00:00:00 +0000

94% · Inference speed 94%, Serving systems 89%

Unlike autoregressive large language models (AR-LLMs), diffusion LLMs (dLLMs) generate text by iteratively denoising many tokens in parallel, offering a promising path to efficient inference. Existing serving frameworks, however, schedule dLLM requests with policies designed for AR models, namely first-come-first-served (FCFS) admission and prefill priority. We find that both policies are ill-suited to dLLMs, wasting much of their speed. FCFS ignores that a dLLM request's decoding cost is highly heterogeneous and unknown at arrival, causing severe head-of-line blocking. Prefill priority, widely used in AR-LLM serving to keep time-to-first-token low, sharply cuts dLLM efficiency, as rounds that run only prompts leave most of the batch idle. We present DWISE, a Denoising Workload-Informed Scheduling Engine for efficient dLLM serving. DWISE schedules requests by their remaining work, measured in a block-level unit shared by prefill and denoising, so both phases share a round. To predict the unknown denoising workload, DWISE combines a lightweight prompt-based predictor at admission with a single refinement from the first denoising round's hidden states, which capture model-specific information absent from the prompt and require no additional model execution. Implemented in SGLang and evaluated on five models from three dLLM families under bursty workloads, DWISE reduces mean end-to-end latency over SGLang by 5.4$\times$ on average and up to 14.9$\times$ while preserving task accuracy. The code will be released publicly.

ConfluxServe: Reconciling Resource Utilization and Parallelism Efficiency across Prefill and Decode in LLM Serving

Wenhao Chen, Hong Zhang — Fri, 30 Oct 2026 00:00:00 +0000

93% · Inference speed 93%, Serving systems 92%

Prefill and decode are complementary in resource usage, but often divergent in parallelism preferences. Executing the two phases separately preserves their preferred parallelism but leaves complementary resources underutilized, while co-batching improves utilization but often requires a uniform parallel configuration that hurts execution efficiency. To resolve this utilization-efficiency dilemma, we propose Conflux Parallelism (XP), a parallel execution abstraction that co-executes prefill and decode with different parallelism within the same batch. Conflux Parallelism exploits operator-level Transformer structure to construct a heterogeneous execution path with different batching and parallelism strategies across operators, combining the best of co-batching and phase-specific execution. Building on Conflux Parallelism, we design ConfluxServe, an SLO-aware serving system that maximizes cluster goodput under TTFT and TBT constraints. Evaluation on production workloads shows that ConfluxServe improves serving goodput by 1.26-1.62x over the strongest baselines.

Low-Bit GDN Recurrent-State Quantization: Global Gradient-aware Allocation and Pure-Write Sliding Windows

Qinghang Huang, Xuchu Huang, Xunzhao Yin — Fri, 30 Oct 2026 00:00:00 +0000

93% · Compression 93%, Inference speed 60%

Linear attention stores historical information in a recurrent state of fixed size, providing a compact memory representation for long-context inference. Reducing state precision further lowers per-request storage requirements, but quantized writeback errors enter subsequent recurrences, degrade prediction quality on long-reasoning tasks, and can be accompanied by repetitive generation and excessively long outputs. We propose Global Gradient-aware Allocation (GGA) and the Pure-Write Sliding Window (PW) to address precision allocation and recent-information maintenance. GGA combines candidate quantization residuals with gradient sensitivity to future prediction loss to allocate head bit widths under a model-wide mixed-precision budget. PW independently maintains recent pure writes as FP8 vector pairs, propagates them through the full GDN state transition, and merges them into the low-bit historical state only after they leave the window. Across seven evaluations on Qwen3.5-4B/9B, GGA allocation alone improves over STEPQuant@4 by 19.48/29.84 percentage points. Adding PW-S8 yields seven-task macro-averages of 69.83%/78.63%, improving over STEPQuant@4 by 38.39/41.21 percentage points, respectively. Both precision reallocation and window maintenance improve task scores while mitigating excessively long outputs: at @6, GGA reduces mean generation length, averaged equally across tasks, by 10.45%/14.65% relative to STEPQuant; adding PW-S8 to GGA@4 further reduces mean generation length by 32.32%/36.77%. Comparisons at similar storage budgets further show that using additional capacity to maintain recent writes yields higher task scores and shorter mean outputs than solely increasing the bit width of the accumulated state.

LinQuant: Functional Invariances for Low-Precision Linear Recurrences

Neehal Tumma, Philipp Nazari, Rohan Cherukuri, Daniela Rus, T. Konstantin Rusch — Fri, 30 Oct 2026 00:00:00 +0000

91% · Compression 91%, Inference speed 56%

Linear recurrences are emerging as practical alternatives to softmax attention in large language models, appearing in both standalone and hybrid architectures. However, despite their rapid emergence, low-precision inference remains much less developed for these models than for softmax attention. During prefill, chunkwise-parallel algorithms cast much of the recurrence as dense matrix multiplications, suggesting a natural path to low-precision execution. Simply quantizing these operations in their native form, however, can be fragile. In this work, we identify functional invariances that preserve the full-precision computation while making the operands more amenable to quantization. Building on this observation, we introduce LinQuant, a framework for low-precision recurrent prefill based on seven such invariances, including rotations, diagonal rescalings, and alternative placements of recurrent decay and write factors. We derive when these transformations remain valid across common recurrence families spanning linear, gated, and delta-rule updates, and show that they provide complementary degrees of freedom for improving low-precision accuracy. Finally, we translate this framework to production by augmenting SSDi8, an existing low-precision prefill pipeline for Mamba-2, with the applicable invariances. LinQuant improves perplexity at matched latency on Mamba-2 1.3B, 2.7B, and 8B and Nemotron-H 8B, demonstrating the practical utility of invariance-based quantization for recurrent models.

DolphinRouting: Cluster-Level Binary Routing for Disaggregated LLM Inference Serving

Yishuo Yuan, Shaoyu Yang, Chunrong Fang — Fri, 30 Oct 2026 00:00:00 +0000

91% · Inference speed 91%, Serving systems 71%

Prefill-decode (PD) disaggregation isolates the two phases of large language model serving, but request-level path selection must account for prefill work, cache reuse, and decode demand. We present DolphinRouting, which learns this choice from end-to-end SLO feedback under a given P/D resource allocation. Requests are clustered by raw prompt length, cache-adjusted input lengths on both worker pools, and output-length information. Offline policy search learns one binary action per cluster. Online feature extraction, cluster assignment, and table lookup select local or remote prefill. We evaluate DolphinRouting on a real PD-disaggregated cluster using three public benchmarks, a mixed workload, and real data. Across the mixed workload and real data, DolphinRouting with $k$-means reduces P90 TTFT by 5.7--7.9\% and P90 TPOT by 2.6--10.1\% relative to Dynamo.

★ Distributed LLM Inference via Communication-Efficient Speculative Decoding

Joohyun Shin, Byeonghyeon Park, Keumseo Ryum, Joonhyuk Kang — Wed, 28 Oct 2026 00:00:00 +0000

96% · Inference speed 96%

Distributed speculative decoding has emerged as a promising approach for device--server collaborative large language model (LLM) inference, but its decoding speedup is often limited by the vocabulary-scale communication overhead. In existing distributed speculative decoding methods, exact rejection sampling requires transmitting full token distributions over the LLM vocabulary between the device and the server, which becomes costly in bandwidth-constrained environments. To address this issue, we propose DISCO, a communication-efficient framework for distributed speculative decoding, that eliminates vocabulary-scale distribution transfer while preserving the exact target distribution. DISCO achieves this by introducing a server-side replica of the draft small language model (SLM), which enables the server to locally obtain the draft-side distribution required for rejection sampling. As a result, DISCO exchanges only drafted token indices and scalar probability values between the device and the server. Experimental results on multiple datasets show that DISCO consistently achieves higher decoding speedup than existing distributed speculative decoding methods, particularly under stringent bandwidth constraints.

★ INC-MoE: Overlapping In-Network Collectives and Expert Computation for Expert-Parallel Mixture-of-Experts Inference

John Stephan, Razvan Puscasu, Xu Yong, Ding Jiutao, Guan Zixuan — Wed, 28 Oct 2026 00:00:00 +0000

96% · Inference speed 96%, Serving systems 89%

In expert-parallel mixture-of-experts (MoE) serving, dispatch moves tokens to the accelerators hosting their selected experts, and combine returns expert outputs to each token's owner. Both operations lie on the critical path and can dominate MoE layer latency in multi-node deployments. We present INC-MoE, an expert-parallel MoE layer that offloads dispatch and combine to the network switch and overlaps both collectives with each other and with expert computation. Dispatch uses in-network AllGather through switch-side multicast, while combine uses in-network ReduceScatter through switch-side reduction. Both collectives use message semantics between the end hosts and the switch. Because the in-network collectives load opposite directions of each link, executing them concurrently reduces bandwidth-bound communication time by $(N-2)/(2(N-1))$ relative to concurrent ring-based collectives: 33% for four ranks and 43% for eight. INC-MoE partitions tokens into bands and processes them through a three-stage AllGather-feed-forward network (FFN)-ReduceScatter pipeline inside a single persistent kernel. At steady state, role-specialized thread blocks gather the next band, compute the current band, and reduce the previous band concurrently, overlapping both collectives with each other and with computation using device-side synchronization. INC-MoE integrates as a drop-in replacement for the MoE layer of vLLM. We evaluate it in an emulated four-node scale-out deployment with a dedicated accelerator emulating an in-network computing (INC) switch. On OLMoE prefill with four-way expert parallelism, INC-MoE reduces time to first token (TTFT) by up to 33% and increases throughput by up to 50% relative to vLLM's default collective path, while matching its outputs.

★ CoS-VQ: Coordinated Shared-Scale Vector Quantization for Large Language Models

Zukang Xu, Zhixiong Zhao, Xing Hu, Zhe Jiang, Xin Si, Dawei Yang — Wed, 28 Oct 2026 00:00:00 +0000

96% · Compression 96%, Inference speed 71%

Large language models (LLMs) require substantial memory, motivating low-bit weight quantization for efficient deployment. Existing vector quantization (VQ) methods often improve accuracy by increasing the effective coding dimension, which can complicate independent lookup and local codeword updates. We propose CoS-VQ (Coordinated Shared-Scale Vector Quantization), which preserves low-dimensional codewords while jointly optimizing hard indices and their shared scale across positions. The key challenge is scale-induced coupling: the optimal scale depends on the entire codeword combination and, in turn, changes candidate rankings across positions, so pointwise updates can stall even after scale re-optimization. CoS-VQ collects candidates at multiple probe scales and constructs a tangent separable upper bound for the residual-aware quadratic objective. By reducing candidate comparisons to lower envelopes of lines over positive scales, it identifies finitely many representative intervals for joint index--scale updates without exhaustively enumerating codeword combinations. On Qwen3.8-27B, CoS-VQ outperforms QTIP and LLVQ on the evaluated language modeling, coding, and reasoning benchmarks. Ablations from the same initialization further show larger gains from joint updates than from scale-only refinement, highlighting joint selection as an alternative to increasing codeword dimensionality. The code is available at https://github.com/xzktx003/Cos_VQ.

★ CRISP: Co-designing Pruning and Scheduling for Efficient MoE Inference on Edge Servers

Congwei Zhang, Youming Tao, Yifei Zou, Yu Yang, Ruirui Zhang, Xiaoyu Zhang — Wed, 28 Oct 2026 00:00:00 +0000

96% · Inference speed 96%, Serving systems 70%, Compression 54%

Mixture-of-Experts (MoE) inference on memory-constrained edge platforms is often dominated by expert-weight movement rather than computation. Although each token activates only a small subset of experts, routing is input-dependent, so all experts must remain accessible and memory movement dominates latency. Existing approaches such as uniform pruning and expert dropping overlook two properties of MoE execution: expert combinations are highly skewed in frequency, and experts contribute asymmetrically within each combination. We present CRISP, a combination-aware framework that co-designs pruning and runtime scheduling for efficient MoE serving on memory-constrained edge platforms. CRISP reorders neurons by importance so one stored expert can support multiple execution widths, assigns differentiated execution to hot, warm, and cold combinations based on frequency and expert role, and exploits heterogeneous widths to overlap expert loading with computation. Across Mixtral-8x7B and DeepSeek-MoE-16B, CRISP improves throughput by up to $2.1\times$ at the same memory budget while limiting accuracy loss to 3.6\% relative to full-width inference. Our results highlight that efficient edge MoE serving requires jointly optimizing which expert capacity is preserved and how expert execution is scheduled.

★ HiLo-RD: Cross-Budget Self-Distillation for Masked Diffusion Language Models

Xiang Li, Chengting Yu, Jungang Li, Yilun Zheng, Junzhuo Li, Xingyao Zhang — Wed, 28 Oct 2026 00:00:00 +0000

95% · Compression 95%, Inference speed 67%

We propose HiLo-RD (High-to-Low Budget Representation Distillation) to improve masked diffusion language models (MDLMs) under a limited generation budget. Unlike autoregressive models, MDLMs choose both what to generate and where to commit it. This creates two coupled challenges: predicting useful content before reasoning is complete, and deciding when to fix answer tokens that shape later predictions. We identify budget-induced premature commitment: even when a correct solution fits, the model can fix answer fragments near the window boundary before resolving intermediate steps. These fragments become context for later predictions and redirect subsequent reasoning. Existing post-training methods do not directly transfer the stronger trajectories that the same model follows with a larger window. HiLo-RD addresses this gap by (1) pairing large-window teacher states with cropped small-window views and (2) matching their hidden representations at committed positions. This supervision guides both token prediction and commitment order without an external teacher. Across mathematics benchmarks, HiLo-RD improves low-budget accuracy and shortens responses; these gains extend to code generation without code-domain training. On GSM8K, accuracy rises by 9.4 percentage points as mean readable length falls by about 25%. Attention diagnostics and fixed-budget order interventions support the proposed mechanism. Together, these results show that large-window reasoning can improve small-window decoding.

Knowledge distillation across teacher architectures for compact diabetic retinopathy grading

Mossa Manisa, Habiba Jahan Priya, A. K. M. Fakhrul Hossain — Wed, 28 Oct 2026 00:00:00 +0000

94% · Compression 94%

Diabetic retinopathy grading from retinal fundus photographs is important but difficult to deploy on resource-constrained hardware. We conducted a controlled study of offline response-based knowledge distillation for five-class diabetic retinopathy grading on APTOS 2019, comparing ResNet-50, MobileNetV3-Large, and Swin Transformer-Tiny (Swin-T) teachers distilled into a MobileNetV3-Small student. A 3 x 4 x 2 grid of teacher, temperature, and hard-label weight was screened at one seed, while six configurations and the no-distillation baseline were evaluated across three seeds. The highest observed single-seed quadratic weighted kappa was 0.9093 (macro-F1 0.6851) for Swin-T at temperature 4 and hard-label weight 0.7; the corresponding three-seed mean was 0.9024, compared with 0.8804 for the baseline. The student used 1.52 M parameters and 5.94 MB, representing an 18.07x parameter reduction relative to Swin-T. Temperature and hard-label weight effects were teacher- and configuration-dependent, while small performance differences were sensitive to random seed. Knowledge distillation produced a compact student with competitive performance under the evaluated protocol, but independent external validation remains necessary.

Let the Gate Decide: Per-Lookup Precision for Quantized Conditional Memory in LLMs

Tianqi Chen, Xunchao Li, Linchuan Xie, Guanghua Yu, Jianchen Zhu, Peisong Wang — Wed, 28 Oct 2026 00:00:00 +0000

94% · Compression 94%, Inference speed 89%

Conditional memory layers, such as DeepSeek's Engram, attach large hashed n-gram tables to a language model, and removing them from DeepSeek-V4.1-Flash costs 36.8 PopQA points. Low-bit quantization loses long-tail facts first, and the standard remedy, giving more bits to frequently accessed rows, backfires: refining lookups by access frequency is worse than refining them at random. We trace this to the gate that scales each retrieved row. A lookup's quantization error is weighted by its gate, which is small for frequent n-grams, large for the rare n-grams of long-tail entities, and computable from a 1-bit code. GateRQ therefore stores each row as a ladder of 1-bit residual codes, keeps level 1 in GPU memory to compute the gate, and reads deeper levels only for the lookups the gate ranks highest. Without retraining, GateRQ cuts GPU table memory by 7.8x on DeepSeek-V4.1-Flash, which raises vLLM throughput by 11%-31% over FP8 tables, while losing about 1 PopQA point at 1.50 bits read per lookup and 0.2 points at 2.10 bits across two models.

How Much Does the Hint Matter? Scope Conditions for Hindsight Self-Distillation in Financial Reasoning

Johnny Yu, Cyril Gorlla — Wed, 28 Oct 2026 00:00:00 +0000

94% · Compression 94%

Financial institutions operating under data-residency constraints are restricted to open-weight models small enough to locally host, which trail frontier systems on multi-step quantitative reasoning. On-policy distillation narrows such gaps by supervising a student on its own rollouts, but traditional OPD requires a strong external teacher, and on-policy self-distillation (OPSD) conditions on a full reference solution. We ask how much of that signal is necessary. We present Fin-SD, a localized hindsight self-distillation recipe for multi-step reasoning. Given a problem the student fails, the student itself identifies the first incorrect reasoning step and writes a short critique of that step. We truncate the student's trajectory immediately before the error (aiming to obtain a correct prefix), condition a frozen copy of the student on the same prefix plus the hint, and distill that hint-conditioned teacher into the student over a short rollout under reverse KL. No external model contributes a gradient. We apply Fin-SD to GPT-OSS-20B on a custom multi-step financial reasoning benchmark that requires deriving formula inputs from realistic scenarios via accounting conventions. We find: (i) replacing the critique with a generic-feedback control matches Fin-SD's entire 3.03 pp accuracy gain over the base model on the FinanceReasoning Hard subset, suggesting the gain comes from where supervision is applied, not hint content; (ii) the model writing the hint does not need to be stronger than the model receiving it: hints from same-family GPT-OSS-120B, at roughly 6x the parameters, yield no materially greater benefits than self-written hints; and (iii) Fin-SD matches on-policy self-distillation from full reference solutions on accuracy while generating 9.5% fewer tokens. With Fin-SD, general reasoning on MMLU-Pro and a fourteen-metric compliance audit on FinTrust show no material regression.

Live-OPSD: On-Policy Self-Distillation for Proactive Streaming Large Video Models

Kaixuan Du, Xin Wan, Dai Guan, Ming Chen, Xiaoxi Jiang, Guanjun Jiang — Wed, 28 Oct 2026 00:00:00 +0000

93% · Compression 93%

Streaming multimodal assistants must make response decisions while events are still unfolding. They need to recognize when the available evidence warrants a response, produce an appropriate continuation, and remain silent otherwise. Supervised fine-tuning learns from recorded assistant turns, but does not directly train on the response prefixes produced by the evolving policy itself. We introduce Live-OPSD, an on-policy self-distillation method that uses privileged hints available only during training to supervise the policy on its own outputs. The policy first generates a continuation from the streaming context alone. The same model then re-evaluates that continuation with a hint describing the relevant evidence, and the resulting distributions guide the unhinted policy. On Qwen3-Omni-30B-A3B, Live-OPSD attains an average score of 59.64 on LiveProBench, 5.26 points above the strongest continued-SFT baseline, and stays competitive on OVO-Bench (65.84). These results indicate that guiding the policy on its own outputs with privileged hints improves both when and how a streaming assistant responds.

Efficient Quantization Training with Stochastic Rounding via Structured Subspace Randomness

Zeyu Wang, Yiheng Duan, Tao Chen, Donghyun Koo, Sihan Liu, Wenbo Zhang — Wed, 28 Oct 2026 00:00:00 +0000

92% · Compression 92%

Low-precision training often relies on stochastic rounding (SR) to keep quantized gradients unbiased and prevent small updates from rounding to zero. Standard SR draws an independent random threshold for every gradient element. However, accelerators generate random numbers on vector units whose throughput lags far behind their matrix units. Drawing one threshold per element saturates the vector unit and can make INT8 training with SR slower than BF16 training. Sharing random thresholds across tensor elements removes this cost but forces those elements to round in the same direction and shifts their rounding errors by a common offset. We find that this offset degrades training primarily through the activation gradient, which propagates the error backward into earlier layers. Because earlier layers cannot cancel an incoming offset using their own local rounding errors, the offset must be removed at the layer where it originates. We propose Structured Subspace Stochastic Rounding (S3R), which quantizes each upstream gradient once for both the activation and weight gradient matrix multiplications and shares a single threshold per token across the hidden dimension. Its error correction, S3R-EC, subtracts the structured mean rounding error from the activation and weight gradients while leaving the optimizer unchanged. In LLM pre-training experiments, uncorrected threshold sharing degrades convergence, whereas S3R-EC eliminates this degradation and matches the validation loss of independent SR. Furthermore, S3R-EC accelerates full INT8 training steps of dense LLMs on TPU v5p by 1.24x to 1.32x over BF16 and by 1.11x over independent SR.

Proposing Better Rollouts for On-Policy Distillation

Sirui Chen, Yunzhe Qi, Haobo Xu, Jiaru Zou, Mengting Ai, Zihao Li — Wed, 28 Oct 2026 00:00:00 +0000

91% · Compression 91%

On-policy distillation (OPD) improves language-model post-training by letting the student learn from its own sampled trajectories under teacher supervision, aligning the training process with inference-time behavior. However, the rollout policy in OPD is usually treated as fixed: each round samples directly from the current student. We show that this choice can create a compounding teacher--student mismatch, because it requires the teacher to provide effective supervision on whichever trajectories the student samples, including ones less reachable by the teacher itself. On such trajectories, teacher supervision can become noisy or even misleading, leading to undesirable student patterns such as repetition. Moreover, since the updated student in turn defines the rollout policy of subsequent rounds, these patterns persist and potentially amplify over training. This failure mode suggests that improving OPD requires treating the rollout policy itself as an optimization target, rather than only modifying the distillation objective. We propose Mixture Proposal On-Policy Distillation (MPOPD), which derives a student-anchored, teacher-guided rollout distribution from a local trust-region objective based on relative teacher support. The resulting target is a geometric mixture of the student and teacher policies, and we introduce a proposal policy that approximates this mixture, enabling rollout generation in a single forward pass. Experiments on 16 math, knowledge, and medical reasoning benchmarks show consistent gains over OPD and recent variants. On math reasoning, our method improves average accuracy by 6.91 over OPD on Qwen3-4B-Base while reducing average generation length by 41%, suggesting that rollout policy design provides an effective way to improve OPD.

TDLM: Adaptive Temporal Integration for Diffusion Language Models

Bohang Sun, Max Zhu, Francesco Caso, Christopher Irwin, Jialin Yu, Pietro Li\`o — Wed, 28 Oct 2026 00:00:00 +0000

91% · Inference speed 91%

Masked diffusion language models repeatedly compute representations for unresolved tokens, but conventional denoisers do not explicitly retain these representations for reuse across steps. We introduce Temporal Self-Conditioning Diffusion Language Model (TDLM), which lets each unresolved position attend to its own history of hidden representations within a frozen denoiser. This gives later predictions access to computation already performed. Training uses self-generated trajectories and a simple cross-entropy objective for tokens revealed between consecutively sampled decoding steps, while causal temporal attention permits parallel evaluation during training. On LLaDA, the temporal modules add only approximately 0.84% trainable parameters. Across LLaDA and Dream, the main comparison shows up to 7.15$\times$ faster generation than the corresponding original diffusion language model, with competitive answer quality on the evaluated tasks. These results motivate temporal memory as a practical way to make more effective use of repeated denoising computation.

Channel Selection for Mixture-of-Experts Decoding Needs No Calibration Data

Geunhyoung Kim, Sung Ju Hwang — Wed, 28 Oct 2026 00:00:00 +0000

91% · Inference speed 91%, Compression 34%

Mixture-of-Experts (MoE) models are bottlenecked by loading routed-expert weights from memory during interactive decoding at small batch sizes. Because most intermediate channels in a routed expert contribute little to any single token, channel selectors predict which channels matter and read only their weights, typically keeping channels above a threshold calibrated on offline data. We show that channel selection needs no calibration data. First, a calibrated threshold does not fix how much a selector reads, because the number of channels passing it varies across inputs; Prox with thresholds calibrated for an 18% read budget actually reads only 12.9% on average across our benchmarks. Second, top-k keeps each token's k highest-scoring channels and needs no calibration data. At the calibrated budget, a selector's top-k version matches or beats its threshold version on the same score in 40 of 43 comparisons at 13% and 18% budgets, including all 23 on generation tasks. The top-k versions also avoid the repetition-loop collapse that calibrated thresholds cause on Gemma 4. Calibrating a threshold thus provides neither of the two things selectors use it for: a known read budget and a consistent accuracy gain at that budget. We furthermore introduce a reader built entirely from the model's weights, which screens channels with a ternary sketch and recovers the skipped ones with a low-rank map. Across three post-trained MoE models, our reader retains 98.1% of full-model accuracy at 17.8% of routed-expert reads, with Triton kernels running MoE blocks 1.60 to 1.90x faster than optimized vLLM at batch one and up to 2.69x at larger batches on the same three models.

Request-Adaptive Execution for Efficient Hybrid LLM Inference

Fahao Chen, Xuzhe Yan, Qi Zhang, Peng Li, Dongxiao Yu, Song Guo — Wed, 28 Oct 2026 00:00:00 +0000

91% · Inference speed 91%, Serving systems 84%

Hybrid large language models (LLMs) have emerged as a promising way to combine the modeling capability of the Attention mechanism with the efficiency of state space models (SSMs). However, most existing hybrid LLMs still inherit a static design philosophy: once the Attention-to-SSM layout is chosen, every request must follow the same hybrid execution. This one-layout-fits-all design is inefficient for real-world serving workloads, where different requests rely on contextual information in different ways. This paper presents Morph, a serving system that co-designs the model architecture and runtime for request-adaptive hybrid LLM inference. Its model architecture pairs Attention and SSM layers within each block and uses prefill-oriented gating to select a request-specific layout that remains fixed throughout decoding. Its serving runtime preserves batching efficiency under divergent layer choices through selective batching and enables correct prefix reuse through hybrid-aware context caching. Experiments on six datasets show that Morph achieves a better quality-efficiency trade-off than existing static hybrid baselines. System-level evaluation further shows that Morph improves the serving throughput by up to 3.87x over baselines while substantially reducing inference latency.

Balanced State Subspaces for Mamba-2 State Compression

Shang Tingyu, Cao Wenfei — Wed, 28 Oct 2026 00:00:00 +0000

90% · Inference speed 90%, Compression 59%

Mamba-2 compresses all context into a fixed-size recurrent state that dominates decoding memory and bandwidth. Existing compressors quantize state values or prune state coordinates, scoring each channel independently in the native basis. We examine trained Mamba-2 states and find that energy is almost rank-one, but the output does not read the energetic directions: energy concentrates in a bias-like direction accumulated by the recurrence, while the content the output reads lives in a low-dimensional subspace that is not aligned with the state's coordinates. Even output-aware scores are computed coordinate by coordinate, so they miss the directions that determine the input--output map. Compression should therefore target the write--read subspace rather than individual coordinates. We propose \method{}, a post-training transform that instantiates balanced truncation with full controllability and observability Gramians and applies the rotation exactly via Mamba-2's scalar-identity transition. Because the useful rank varies widely across layers, we further introduce HARP, which allocates ranks per layer from the Hankel spectrum and protects retrieval-sensitive layers, improving both perplexity and retrieval over uniform budgets. Calibration takes a single forward pass, an order of magnitude faster than GHOST's released protocol, with lower peak memory. With HARP, \method{} halves the state dimension at an average WikiText-2 perplexity increase of $0.3$ points ($1.8\%$) across 130M--2.7B, versus $1.6$ ($9.3\%$) for GHOST, and the gap widens at 25\% retention.

★ Vision Token Pruning via Query-Vision Interaction Decomposition

Harshithanjani Athi, Sravan Kumar Ankireddy, Jianzhong Charlie Zhang, Hyeji Kim — Tue, 27 Oct 2026 00:00:00 +0000

96% · Compression 96%, Inference speed 79%

Large vision-language models (VLMs), which process both visual inputs and text queries, incur high inference costs due to the large number of visual tokens, making visual token pruning a natural approach for improving efficiency. Recent work has leveraged query information to guide token selection, often through heuristics derived from attention patterns or token dynamics. However, these approaches do not explicitly exploit the underlying structure of query-vision interactions. We address this gap by observing that the query-vision interaction matrix, naturally induced by the query and key projections already present in transformer attention, is effectively low-rank, with a small number of dominant latent interaction modes capturing query-relevant semantics. Motivated by this observation, we propose Query-Vision Decomposition (QViD), a novel training-free query-aware visual token pruning method that exploits this structure. Extensive experiments on both image and video understanding benchmarks show that QViD consistently improves performance under matched token budgets, with the largest gains appearing under aggressive compression regimes.

★ One Call Ahead: Packed Self-Speculative Decoding with Diffusion Drafting

Yuxiao Chen, Changxiao Cai — Tue, 27 Oct 2026 00:00:00 +0000

96% · Inference speed 96%

Autoregressive (AR) language models generate tokens sequentially, so generating long outputs is costly. Speculative decoding mitigates this by using a lightweight drafter to propose multiple tokens for the target AR model to verify in parallel, while a diffusion-based drafter can produce an entire draft block in one forward pass. In models with AR and diffusion modes, one network both drafts and verifies, but drafting still waits for verification, so the standard schedule needs two sequential model calls per iteration. Existing one-call schedules avoid the wait but incur input size quadratic in the speculative width. Motivated by this efficiency bottleneck, we introduce Packed-SS, a self-speculative decoding schedule that verifies and drafts in a single model call with input size linear in the speculative width, while preserving the target model's distribution. The key idea is to let the diffusion drafter run one call ahead: while verifying the current draft, the model simultaneously constructs the next draft from an older verified prefix and keeps its unresolved suffix for the next iteration. Under a mild monotonicity assumption on the drafter, we further show that Packed-SS requires no more model calls than the standard two-call decoding in the greedy setting. Experiments on Nemotron 3B/8B and SetDLM 1.7B show higher throughput and lower mean wall time than both baselines under greedy decoding.

★ ZipFusion: Fusing Lossless KV Decompression with Attention for Efficient LLM Decoding

Yuetao Chen, Peng Wang, Hong Xu — Tue, 27 Oct 2026 00:00:00 +0000

96% · Inference speed 96%, Serving systems 77%

The key-value (KV) cache places growing demands on GPU memory capacity and bandwidth as large language models serve longer contexts and more concurrent requests. Its footprint limits concurrency, while repeated KV reads during attention can make decoding memory-bandwidth bound. Prior lossless KV cache compression methods reduce transfers between devices but restore the uncompressed KV cache in GPU memory before attention. They therefore leave these repeated KV reads unchanged and require additional work and memory for decompression. We present ZipFusion, which enables attention to consume the losslessly compressed KV cache directly. ZipFusion co-designs the compression format and attention kernel to reduce the bytes read from GPU memory during attention. The kernel reconstructs the original values directly in registers using simple bit operations, avoiding a separate decompression pass and easing the memory-bandwidth bound. ZipFusion also reassigns memory freed by compression across layers to store additional KV cache. Its memory manager reserves space for ongoing generation and makes the remaining capacity available to new requests. We implement ZipFusion on top of FlashInfer and integrate it with SGLang. On an NVIDIA H100, ZipFusion achieves about 1.25x speedup over FlashInfer in decode attention. With 8K context, it improves model decode throughput by up to 26.3% over uncompressed SGLang under the same GPU memory budget. At each method's largest supported batch, it also delivers up to 2.76x the throughput of the best-performing lossless KV cache compression baseline.

★ SPARTA: A Sparse Ternary LLM Accelerator with Adaptive Attention and Unified Datapath for Long-Context Edge Inference

Yanzhen Zhou, Zhengyu Han, Qunyou Liu, Zheyu Yan, Weihong Xu, Cheng Zhuo — Tue, 27 Oct 2026 00:00:00 +0000

96% · Inference speed 96%, Serving systems 31%

LLMs on edge devices are constrained by limited memory and bandwidth. Although 1.58-bit ternary weights greatly reduce model storage, long-context inference is increasingly bottlenecked by KV-cache traffic and attention. Existing accelerators also use separate datapaths for ternary linear layers and attention, incurring hardware and execution overhead. We present SPARTA, a sparse ternary LLM accelerator for efficient long-context edge inference. SPARTA integrates: (1) a phase-aware sparse dataflow exploiting heterogeneous sparsity across compute-bound prefill and memory-bound decode; (2) a unified compute fabric mapping ternary SpMM/SpMV and bit-serial attention onto shared processing elements to exploit zero-weight and zero-bit sparsity; and (3) an adaptive two-stage attention engine combining page-level Top-k selection to reduce K-cache accesses with token-level Top-p pruning to reduce V-cache traffic. SPARTA reduces KV traffic by 4.1-4.8x over full attention and 1.3-1.5x over a state-of-the-art sparse-attention scheme, with task-dependent quality loss. Compared with NVIDIA Jetson AGX Orin, SPARTA achieves 3.9x higher generation throughput and 8.3x higher end-to-end energy efficiency on average; versus AMD Ryzen 7 9700X, the gains are 5.0x and 81.7x. It also delivers 1.6-12.0x higher throughput, 2.0-4.2x higher energy efficiency, and 1.3-51.4x higher area efficiency than FIGLUT, TeLLMe, and MCBP.

★ MCF: A Memory- and Communication-Aware Co-Designed Framework for Long-Context LLM Inference

Tue, 27 Oct 2026 00:00:00 +0000

95% · Inference speed 95%, Serving systems 89%

This paper presents Memory- and Communication-Aware Framework for Long-Context LLM Inference (MCF) that optimizes long-context inference through communication-aware and stage-adaptive KV-cache layout transitions. At the core of MCF is {Memory-Aware Replication and Scheduling} (MARS), a dynamic orchestration mechanism for chunked prefill. MARS exploits early-stage memory headroom by selectively replicating highly reused KV blocks, improving locality and hiding KV propagation latency behind attention computation. As prefill progresses and memory pressure increases, MARS progressively drains redundant replicas and converges the KV cache into a sharded decode-ready layout, avoiding expensive layout conversion at the prefill--decode boundary. For decoding, MCF adopts a query-centric execution model in which KV shards remain stationary while lightweight query tensors are routed across workers. This changes communication from KV-cache-dependent transfer to query-dependent transfer, reducing communication growth with context length. MCF further integrates a hardware-software co-designed execution layer based on an FPGA compute-NIC datapath, enabling metadata-driven KV routing and collective operations directly in the network path. Experiments show that MCF improves TTFT, throughput, and scalability over state-of-the-art distributed inference systems, with larger benefits as context length and cluster size increase.

DualPace: Decoupled Scheduling and Cooperative GPU--CPU Decoding for LLM Serving

YuXuan Lin, KeQiang Wu, Kingsum Chow — Tue, 27 Oct 2026 00:00:00 +0000

94% · Inference speed 94%, Serving systems 94%

Large language model (LLM) serving relies on batching to improve inference throughput, but the key-value (KV) cache grows with sequence length and limits GPU-resident concurrency. CPU offloading can expand KV capacity using host memory, yet existing systems remain largely GPU-gated: CPU-resident requests make progress only when the GPU scheduler exposes suitable overlap opportunities. As a result, host memory and CPU compute cannot act as an independent serving resource, leading to bursty CPU-side progress and inflated tail inter-token latency (ITL) under load. We present DualPace, an LLM serving runtime that turns the host into an independently scheduled decode engine. DualPace separates scheduling authority across two engines: the GPU engine executes prefills and GPU-resident decodes, while an AMX-enabled CPU engine independently batches and advances CPU-resident decodes. The engines are decoupled in scheduling but cooperative in execution: the CPU can execute a complete decode iteration locally or selectively invoke GPU Dense Assist for dense stages, while offline profiles guide hardware-specific batching and routing. Our evaluations show that, in decode-oriented scenarios, DualPace improves serving throughput by 6.9--43.2\% over NEO, a state-of-the-art open-source CPU-offloading system, and by 11.5--90.6\% over vLLM, a widely used GPU serving system, while maintaining the same level of latency. DualPace also preserves NEO's P99 time-to-first-token (TTFT) advantage over vLLM while sustaining 7.3--50.5\% higher request rates than NEO at the same level of P99 ITL.

BellowsKV: One Latent for Many Layers, Expanded on Demand

Jiamu Zhang, Hanjie Chen — Tue, 27 Oct 2026 00:00:00 +0000

94% · Inference speed 94%, Compression 73%

Large language models (LLMs) now serve as the backbone of agents that reason over many turns, call tools, and pursue long-horizon tasks. To avoid recomputation, a model stores the attention keys and values of all past tokens in a key-value (KV) cache. Each layer maintains its own cache, so KV-cache memory can become the main bottleneck for GPU serving. Most compression methods operate within a single layer, by quantizing the cache entries, reducing the channel dimension, or evicting less important tokens. All three lose substantial accuracy at high compression ratios. Cross-layer methods share one compressed cache across adjacent layers to go beyond this limit. However, these methods also degrade sharply at high compression ratios. We find that the KV-cache representations of adjacent layers are quite different, whereas the input hidden states of adjacent layers are highly similar. We also find that, under the same memory budget, reducing the dimension of this shared latent loses far more accuracy than quantizing it. Based on these findings, we propose BellowsKV, a training-free method that maps the hidden states of n adjacent layers to one shared latent per token. At inference time, each layer reads its own keys and values from this latent through a layer-specific projection. We keep the latent at full dimension and compress it in two other ways: BellowsKV-Quant stores the latent at low precision, which is nearly lossless because the latent has no dominant outlier channels; BellowsKV-Evict makes one shared eviction decision per block, because the layers rank token importance very similarly. Across Llama-3.1-8B, Mistral-7B and Qwen2.5-7B, BellowsKV-Quant retains about 99% of full-cache accuracy at 3.6x compression and 97-98% at 13-14x, where the strongest cross-layer method retains at most 84% at 16x. Code: https://github.com/MorrisZJ/BellowsKV

Steering the Teacher: Improving On-Policy Self-Distillation for Thinking Models

Tiejin Chen, Xiaoou Liu, Hanyang Chen, Hua Wei — Tue, 27 Oct 2026 00:00:00 +0000

93% · Compression 93%

On-policy self-distillation (OPSD) trains a student on its own reasoning trajectories using token-level supervision from a reference-conditioned teacher initialized from the same model. However, access to the reference can induce teacher behaviors that undermine independent reasoning in thinking models, leading to reference hallucinations, where the student cites an unavailable reference to justify reasoning steps, and reconsideration suppression, where the teacher discourages verification and revision at uncertain steps. We introduce Steered-Teacher OPSD (ST-OPSD) to jointly address these failure modes through activation steering. ST-OPSD extracts two directions from model-generated reasoning trajectories and applies them to the teacher's hidden states during distillation target computation, suppressing reference-invoking behavior and encouraging reconsideration with marker-family calibration. The student learns from the resulting teacher distributions without being steered. Experiments on AIME and HMMT show that ST-OPSD improves accuracy over vanilla OPSD and other interventions while restoring reconsideration and reference-hallucination rates to base-model levels.

DAC-Loop: Distillation-Aware Compression for Remote Sensing CNN

Tue, 27 Oct 2026 00:00:00 +0000

93% · Compression 93%

While large-scale models achieve strong results in remote sensing scene classification, deploying them on edge devices remains difficult due to limited compute, memory, and energy budgets. Compact CNN backbones are therefore a practical choice in many real-world settings, offering efficient inference with well-optimized implementations on embedded hardware. To push this efficiency further, we investigate how far a knowledge-distilled student model can be compressed while preserving high accuracy. We propose a Distillation-Aware Compression Loop (DAC-Loop) that integrates (i) knowledge distillation applied both before and after structured channel pruning, and (ii) an iterative compression schedule that alternates structured pruning with unstructured sparsification. Across EuroSAT and NWPU-RESISC45, the DAC-Loop consistently improves the accuracy-compression trade-off relative to standard one-shot pruning and single-stage KD recovery, maintaining strong accuracy even beyond 90% parameter reduction. These results demonstrate that combining distillation and pruning in a unified iterative loop offers an effective path toward compact remote sensing models.

Prefix Keys and Values Matter Late: RIPE Caching for Self-Speculative Decoding in Looped Language Models

Dezhen Lu, Runguo Li, Zixiang Ni — Tue, 27 Oct 2026 00:00:00 +0000

92% · Inference speed 92%, Compression 56%

A looped language model runs one block of layers $R$ times per token, so its key-value (KV) cache holds $R$ planes per recurrent layer, and self-speculative decoding (draft with a few loops, verify with all $R$) reads them all. We find that prefix keys and values matter late: at equal perturbation size, loops 25-32 are $e^{15.0}$ (Huginn-0125) and $e^{12.8}$ (Raven-Llama) times more sensitive than loops 2-9, and the planes approach the last one geometrically, at 0.88 and 0.81 per loop. Whether the deepest plane alone suffices then depends on how ripe the late planes are: on Raven, storing only it is indistinguishable from a bf16-to-fp32 change; on Huginn, 1% of positions carry 74% of its error. We turn this into RIPE, which stores the deepest plane everywhere except a window of recent positions kept at full depth. On Huginn, a 64-position window removes 89% of the tail positions at 11.9% of the full cache's memory (1024-token context), and any window at least as long as the draft block makes batched verification block-independent by construction. Nine pre-registered GPU batches back these results. Simple models of acceptance, decoding cost and memory, two of them checked on held-out measurements, account for the measured self-speculative speed-ups ($2.09\times$ with the full cache, $2.58\times$ with the deepest plane) and decompose the $2.24\times$ fixed-memory throughput gain of one shared plane.

Transporting Quantiles to Codebook: Scalable Vector Quantization without Codebook Collapse

Yuwei Zeng, Zekun Shi — Tue, 27 Oct 2026 00:00:00 +0000

92% · Compression 92%

Effective discrete representation learning with VQ-VAE serves as a fundamental component of modern autoregressive generative models. However, it is widely known to suffer from codebook collapse and inefficient code utilization. To address this issue, we propose FlowVQ, a simple yet effective alternative to standard vector quantization by formulating neural quantization as a joint modeling and clustering process. Specifically, we learn the latent space using a lightweight flow matching model alongside training, corresponding to a probability flow ODE that transports an often fixed well-behaved source distribution while preserving probability mass. Based on this, we define the codebook C_1 by transporting even-quantile partitioned prior codes C_0 from the source space through the learned flow, and obtaining the Voronoi centroids under the induced distribution, leading to a structured discretization. We evaluated on image tokenization and generation, as well as on two tasks requiring high-precision tokenization, demonstrating scalability and improved expressivity of the codebook with full utilization.

INTRA: Interleaved Non-contiguous Token spaRse Attention

Yuxuan Xia, Wenbo Lu, Shaoyi Zheng, Shengjie Wang — Tue, 27 Oct 2026 00:00:00 +0000

91% · Inference speed 91%, Serving systems 36%

Transformers achieve strong performance across modalities but are bottlenecked by the quadratic cost of attention. Sparse attention can reduce this cost, but effective sparsity must preserve useful long-range communication and map well to GPU hardware. We introduce INTRA--Interleaved Non-contiguous Token spaRse Attention--a hardware-aware static sparse attention design based on a relay-communication view under Computational Query Set (CQS) constraints. Rather than directly covering all token-to-token interactions within each sparse layer, INTRA uses complementary sparse groups to provide short mediated communication paths across layers. Scatter forms relay groups that sample broadly across local neighborhoods, while Gather reconnects these groups in subsequent layers. Under a partition model, this structure guides CQS-compatible sparse patterns with global reachability. Our CQS-aware FlashAttention-style kernel computes structured indices at runtime, loads selected non-contiguous tokens, and packs them densely for computation. On an A100, INTRA achieves a $1.53\times$ speedup in $2K$ FLUX.1-dev generation, with better measured image quality than evaluated CLEAR configurations after LoRA self-distillation. On LLaMA-3.1-8B, it improves long-context prefill efficiency with strong teacher-forced retrieval performance under matched LoRA adaptation. Attention-only benchmarks further show up to $8.4\times$ and $10.4\times$ speedups over MoBA and density-constrained NSA, respectively.

★ KVStream: Adaptive Parallel Storage Management for KV Caching in Large-Scale LLM Serving

Mon, 26 Oct 2026 00:00:00 +0000

95% · Inference speed 95%, Serving systems 85%

Long-context LLM serving for agentic, coding, and document workloads reuses prefixes whose key-value (KV) cache outgrows GPU and host memory, putting storage I/O on the critical path of the inference step for retrieving prefixes. A single node-local NVMe absorbs all retrievals, while the parallel file system or cloud file store that nodes already mount carries almost no inference traffic. We cast prefix retrieval as aggregate-bandwidth scheduling across these tiers and present KVStream, a KV cache engine that batches a step's reads, defers writes behind the forward pass, replicates each chunk on every tier, and lets each tier adaptively claim the next unread chunks once it is free. Our performance model shows this routing attains the aggregate-bandwidth bound without prediction or tuning, whereas any static split misses it by a margin that grows as per-tier bandwidths drift or fluctuate. It also predicts that the time-to-first-token (TTFT) gain grows with load, as faster retrieval drains the queue, and disappears when host memory or arrival order keeps reuse in cache. Plugged into LMCache and evaluated on four models, six datasets, and up to 128 GPUs, KVStream retrieves within 5% of the bound, delivers up to 1.49$\times$ higher offline throughput than LMCache, and lowers mean TTFT by up to 6.1$\times$ under bursty arrivals.

★ TranSAGE: Automating Spatial Accelerator Generation for Large-Scale Transformer Models

Chi Zhang, Yi Sun, Weichuang Zhang, Quan Chen, Jieru Zhao — Mon, 26 Oct 2026 00:00:00 +0000

95% · Inference speed 95%

Large-scale Transformers have achieved remarkable success across NLP, computer vision, and robotics, but their growing computational and memory demands challenge efficient inference. Spatial accelerators offer advantages in on-chip bandwidth and energy efficiency, yet designing them for large models remains costly and time-consuming. Existing accelerator generators typically focus on single-operation support, overlooking the benefits of multi-operation pipelining, which limits throughput and energy efficiency. We propose TranSAGE, a compiler-based framework that jointly optimizes scheduling for large-scale Transformers under multi-operation pipelining and generates corresponding accelerator implementations. Our approach integrates a design space exploration (DSE) engine with multi-operation dataflow optimization and a dynamic programming-based graph partitioning algorithm to efficiently identify pipeline-friendly operation clusters and their optimal dataflow parameters. We further design spatial architecture primitives and lowering passes to automatically generate efficient RTL implementations. Evaluations on four representative LLMs demonstrate that TranSAGE achieves up to 2.41x higher throughput and 13.04x higher energy efficiency than commercial GPU/NPU baselines. Compared with state-of-the-art accelerator generators, TranSAGE achieves up to 11.33x higher throughput and 6.07x higher energy efficiency. Our code is available at https://github.com/sjtu-zhao-lab/TranSAGE.

Learning Structured Sparsity for Large Models: Bridging Model Compressibility and Efficient Inference

Tianxiang Xia, Kai Lion, Niao He, Michael Muehlebach, Melanie Zeilinger, Hao Ma — Mon, 26 Oct 2026 00:00:00 +0000

94% · Inference speed 94%, Compression 62%

Scaling laws continue to reward increases in model scale, while the resulting computational and memory demands drive greater reliance on cloud infrastructure for inference. Yet unreliable network connectivity, communication delays, and privacy requirements can make such reliance impractical, motivating local execution on edge devices. Local deployment requires models to fit within limited memory and execute efficiently with the available computational resources. Recent work has shown that learning sparse-plus-low-rank representations enables elastic compression of a single checkpoint to multiple memory budgets. However, fewer parameters do not necessarily translate into faster inference. Indexing overhead and irregular memory access in unstructured sparse components can offset computational savings. To address this critical deployment barrier, we introduce S-SALAAD, a unified framework that jointly learns the model weights and their low-rank and structured sparse components during training. The framework supports a family of hardware-compatible structured sparsity patterns. It couples this structure learning with budget-aware compression and dedicated CPU and GPU kernels that exploit the learned structure for acceleration, enabling efficient inference from a single checkpoint across memory budgets. Experiments on Llama2 and Qwen3 models, covering both pretraining and distillation, show that S-SALAAD advances the empirical Pareto frontier in perplexity, memory, and inference latency across multiple hardware platforms relative to the evaluated baselines. These results show that structured learning can translate elastic compressibility into reduced memory requirements and lower inference latency, enabling practical deployment of large models on edge devices.

Efficient Omni-modal Large Language Model Inference via Per-Query Modality Selection

Shuqi Liu, Yuzhou Cao, Luke Ong — Mon, 26 Oct 2026 00:00:00 +0000

92% · Inference speed 92%

Omni-modal large language models (Omni-LLMs) can jointly process a broad range of modalities, such as text, audio, and video, enabling stronger multimodal understanding and reasoning capabilities, but often at substantial inference cost. We observe that the benefit of additional modalities varies across queries, indicating that a fixed modality configuration can be inefficient. Motivated by this, we study query-wise modality selection, which keeps the underlying Omni-LLM fixed while adaptively selecting which modalities to provide for each query. We formulate the problem as an accuracy-cost optimization problem and develop a cost-sensitive learning approach for modality selection. We evaluate our approach on WorldSense using Qwen2.5-Omni and Qwen3-Omni under three inference-cost measures: input tokens, latency, and monetary cost. Across both models, query-wise modality selection consistently improves the accuracy-cost trade-off and can even outperform the standard setting in which all available modalities are provided to the Omni-LLM. For Qwen2.5-Omni, our cost-sensitive approach matches or exceeds this full-modality accuracy while requiring only approximately 41% of the input tokens, 39% of the latency, and 61% of the monetary cost.

★ Spike-Aware Gradient-Based Mixed-Precision Quantization of Spiking Neural Networks for Vision Applications

Rajshekhar Rakshit, Devgokul B. V, Rakshitha Kalkura, Gopalakrishnan Srinivasan — Fri, 23 Oct 2026 00:00:00 +0000

96% · Compression 96%

In recent years, Spiking Neural Networks (SNNs) have gained massive attention as an energy-efficient neural computing paradigm due to their event-driven and sparse computing capability. However, maintaining accuracy under low-precision constraints, necessary for practical deployment, has been largely unexplored for SNN models. In addition, existing Hessian-based approaches for quantization, such as HAWQ-V2, incur significant memory and computational overhead due to temporal unrolling across multiple SNN timesteps. We propose a spike-aware gradient-based mixed-precision quantization framework for SNNs that adaptively assigns bit-widths using first-order gradient information, avoiding costly second-order Hessian computations. We evaluate our method across various SNN-based architectures, including VGG, ResNet, and Vision Transformer models obtained from both ANN-to-SNN conversion and direct training frameworks, on CIFAR10/100, ImageNet, and DVS-CIFAR-10 datasets. Notably, on ResNet-34 evaluated on ImageNet, our method achieves 73.30% accuracy at 4.78$\times$ compression, which is comparable to that provided by HAWQ-V2, while reducing memory consumption by 2.31$\times$ and bit-allocation runtime by 100$\times$. We also present an energy analysis for mixed-precision SNNs showing that the proposed quantization strategy enables energy-efficient deployment without compromising performance.

★ Extreme Compression of Per-Layer Embeddings via Cross-Layer Latent Vector Quantization

Jiacheng Liu, Yaxin Luo, Xiaohan Zhao, Peng Sun, Xinyi Shang, Jiacheng Cui — Fri, 23 Oct 2026 00:00:00 +0000

96% · Compression 96%, Inference speed 85%

Scaling embedding capacity alongside Transformer computation is an emerging trend in language models. This trend makes high compression rates increasingly useful. Hence, we study extreme compression of Per-Layer Embeddings (PLE), which follow this trend and account for 35.2% of the parameters of Gemma 4 E4B. At such rates, conventional quantizers give each token-layer vector its own code of only a few bytes. We instead develop OPTIC (Output-Path-fitted, Token-shared Indexed Codes), which stores one quantized code tuple per token and decodes it at every layer with a layer-specific affine decoder. Codes and decoders are fitted to the PLE output path rather than to the stored vectors. At about $100\times$ PLE compression of bfloat16 (BF16) E4B, OPTIC retains an aggregate 97.8% of the uncompressed model's performance across seven benchmarks, whereas product quantization, the strongest of six conventional baselines, retains 66.3%. At $594\times$ and $1,002\times$ PLE compression of BF16 E4B, OPTIC retains 90.5% and 84.1%. On the official E4B release with 4-bit linear weights and 16-bit activations (W4A16), OPTIC retains 96.9% at about $100\times$ PLE compression. During inference at about $100\times$ PLE compression, BF16 E4B uses 29.7% less GPU memory than with uncompressed PLE at the same key-value cache size, while running at most 0.40% slower.

One Teacher, Many Views: Reducing Reference Sensitivity in On-Policy Self-Distillation

Khouloud Saadi, Di Wang — Fri, 23 Oct 2026 00:00:00 +0000

94% · Compression 94%

On-policy self-distillation (OPSD) improves reasoning by training a model on its own trajectories with dense token-level supervision from a privileged copy conditioned on a reference solution. Existing methods typically use a single teacher distribution conditioned on the complete reference. We identify a limitation of this formulation: a reference is only one realization of the reasoning process. Although it provides evidence about the answer, it also encodes an ordering, phrasing, and solution path. Full-reference conditioning can therefore produce correct targets that depend excessively on information unavailable to the student at inference. We show that replacing one valid reference with another, while holding the problem and student state fixed, can substantially alter the teacher distribution, a phenomenon we call reference-induced target instability. To address this instability, we introduce Monte Carlo Privileged Distillation (MCPD), which models privileged context as a distribution over partial reference views and distills the marginal teacher distribution. MCPD evaluates the teacher under multiple views and averages the resulting next-token distributions. We show that this probability-space average is the exact target induced by minimizing the expected teacher-to-student KL divergence over privileged views. Across mathematical and scientific reasoning benchmarks, we show that MCPD produces more reference-consistent supervision and consistently improves over full-reference OPSD.

Low-Latency MoE Inference in Space Data Centers: A Wireless-Power-Transfer Perspective

Haotian Zheng, Zhanwei Wang, Kaibin Huang — Fri, 23 Oct 2026 00:00:00 +0000

94% · Inference speed 94%

Space data centers (SpaceDCs) are emerging as platforms for large language models (LLMs), supported by continuous solar-energy harvesting and high-speed interconnections. The mixture-of-experts (MoE) architecture supports distributed LLM inference by deploying experts across satellites. However, skewed expert activation creates uneven computation power demands despite uniform solar-power supply at the dawn-dusk sun-synchronous orbit (SSO). Consequently, active experts remain power-constrained while harvested power at idle experts is underused. These reduce inference efficiency and increase latency. To address this supply-demand mismatch, we propose the SpaceWPT framework, which uses inter-satellite laser-based wireless power transfer (WPT) to redistribute harvested power. The key design challenge is to balance the computation demands of power-donating and power-receiving satellites under WPT losses: transferring more power accelerates computation at the receiver but leaves less power for the donor. After deriving design principles that simplify WPT control, we characterize the resulting latency tradeoff through a tractable surrogate of average inference latency. Minimizing the surrogate yields a novel two-sided water-filling strategy that determines whether each active expert donates, retains, or receives power according to the water level and its WPT link efficiency. Simulations with Mixtral-8x7B and Phi-3.5-MoE show up to 29.25% lower inference latency than without WPT, with larger gains under more compact formations and higher transfer efficiencies.

When to Follow the Teacher: Advantage-Based On-Policy Distillation

Mahdi Sabbaghi, George Pappas, Hamed Hassani, Adel Javanmard — Fri, 23 Oct 2026 00:00:00 +0000

94% · Compression 94%

Online Policy Distillation (OPD) often outperforms offline methods such as Supervised Fine-Tuning (SFT) by training on trajectories sampled directly from the student. However, standard OPD trains the student to follow the teacher at every state of the trajectory, assuming that teacher guidance is always beneficial. We instead distinguish between what the teacher prefers and what actually helps the current student succeed. We introduce Advantage-Based On-Policy Distillation (AOPD), which weights teacher actions by their counterfactual advantage under the student. Teacher actions that improve the student's expected outcome are reinforced, while those that do not are downweighted. Directly computing these advantages requires trying alternative teacher actions at each step and running new rollouts from each. Instead, we show that the AOPD objective can be estimated using only standard student rollouts and their terminal rewards. Our estimator uses teacher importance weights on the student-sampled actions to provide unbiased advantage-weighted supervision, without any additional rollouts or reward evaluations. This allows AOPD to focus teacher supervision on actions that help the current student while preserving the rollout efficiency of standard OPD. Experiments on math, coding, and Medqa show gains over distillation and RL baselines in both external-teacher and self-distillation settings.

Edit, Don't Imitate: Minimal-Edit On-Policy Distillation for Multimodal Reasoning

Wen Ye, Ailin Deng, Zirui Wang, Shixing Yu, Han Zhang, Tian Yun — Fri, 23 Oct 2026 00:00:00 +0000

93% · Compression 93%

Visual mathematical reasoning can fail before reasoning begins: when a vision-language model (VLM) misreads a diagram, subsequent reasoning proceeds from false premises. This perceptual bottleneck is especially pronounced in smaller models, while accurate textual descriptions can substantially narrow the gap to larger models. Because obtaining such descriptions at inference requires additional annotation or model calls, we propose minimal-edit corrective on-policy distillation. Using reference figure descriptions and verified reference solutions as training-only information, a teacher scores incorrect student rollouts under a minimal-edit objective, and only teacher-disfavored tokens contribute to the corrective loss. Correct rollouts receive task-reward reinforcement learning. This preserves on-policy supervision while avoiding full-response imitation. Across model scales, the method requires no descriptions at test time and meets or exceeds the corresponding base model supplied with a reference figure description at inference. It also improves perception-oriented performance, particularly low-level geometric perception, and transfers to external mathematics and perception benchmarks. Because the minimal-edit teacher evaluates corrections rather than generating its own solution, the method remains effective with a frozen same-size teacher, although its gains are smaller than with a larger teacher on most metrics.

FIGFLEX: A Mixed-Precision FP-INT GEMM Accelerator Co-Designed with Low-Bit LLM Quantization

Fri, 23 Oct 2026 00:00:00 +0000

93% · Inference speed 93%, Serving systems 30%

Weight-only quantization keeps LLM activations in floating point and stores weights as low-bit integers, turning the dominant workload into floating-point-integer (FP-INT) general matrix multiplication. Recent quantizers reach a 2-3-bit average with mixed precision, but hardware has not kept pace: GPUs dequantize weights to FP16 before the tensor-core multiply-accumulate, and bit-parallel FP-INT accelerators run at a single weight precision, so fewer bits save memory but not computation. In this paper, we present FIGFLEX, an algorithm-hardware co-designed mixed-precision accelerator whose bit-parallel datapath scales throughput as precision falls. FIGFLEX contributes (1) a shift-add, radix-4 reformulation of FP-INT multiplication that fuses the multiply-accumulate into a unified, precision-flexible partial-product reduction and roughly halves the partial products, together with a novel INT3 optimization that reclaims the idle datapath radix-4 leaves at that precision; and (2) a mixed-precision quantizer that assigns each channel group its bit-width according to the error GPTQ cannot compensate, with a per-input-channel column scaling whose power-of-two normalization keeps the scaled activations out of the FP16 subnormal range, so that the quantized model maps directly onto FIGFLEX's datapath with the same accuracy as on a GPU. On LLaMA-2-7B, FIGFLEX attains 5.85 and 8.77 WikiText-2 perplexity at 3- and 2-bit average precision, lower than GPTQ, AWQ, SliM-LLM, AMQ, and QuIP# (without fine-tuning), while delivering 1.56x and 1.42x higher whole-chip throughput per area on the linear GEMMs of a 3-bit mixed-precision model than AxCore and FIGLUT, respectively.

SQuAT: Spherical Lattice Vector Quantization-Aware Training

Fri, 23 Oct 2026 00:00:00 +0000

93% · Compression 93%, Inference speed 37%

Vector quantization (VQ), in particular lattice-based methods, offer strong rate-distortion trade-offs for low-bit large language models (LLMs), yet their behavior under quantization-aware training (VQ-QAT) remains poorly understood. We scale lattice VQ-QAT to billion-parameter LLMs and multi-billion-token training regimes, and introduce SQuAT, a shape-gain VQ-QAT approach trained with a separate straight-through estimator (STE) for the shape and gain. SQuAT decomposes weights into direction and magnitude: normalized directions are quantized using an $E_8$-derived spherical codebook, while gains are quantized independently with low-bit scalar quantizers. We show that, unlike a standard STE applied to the full quantizer, which ignores the shape-gain structure, these separate STEs induce tangent-space updates for directions, yielding an implicit form of spherical (Riemannian) optimization. To enable scalability, we develop a distributed exact nearest-neighbor assignment scheme, in which workers quantize disjoint weight blocks and communicate only discrete indices, reducing VQ-QAT training time by 12.8$\times$ and allowing us to scale it to 10B tokens. Across instruction-tuned LLaMA-3.2-1B/3B, Qwen3-4B, and base Qwen3-1.7B, SQuAT consistently outperforms standard VQ-QAT STE under identical quantizers and initialization, and achieves SOTA VQ-QAT accuracy.

Repair Changes Which Layers to Remove: Exact Optima for Depth Pruning of Language Models

Jianchang Su, Dawei Xiang, Yifan Zhang, Wei Zhang — Fri, 23 Oct 2026 00:00:00 +0000

91% · Compression 91%, Inference speed 80%

Depth pruning removes entire transformer layers to make large language models faster, and recent methods recover quality by replacing the removed layers with a training-free linear map. Which layers to remove, however, is chosen with heuristic scores whose distance from the best choice is unknown. We evaluate every subset of layer groups exactly (about 53,000 pruned models across four families, with and without repair) and find that the best layers to remove depend on the repair: linear repair absorbs the interaction among nearby layers and turns the optimum from a split layout into a contiguous block, so methods that choose layers without the repair miss the repaired optimum by up to 189%. We therefore select layers under the repair that will be applied. A greedy search at the cost of SLEB lowers perplexity by 20% to 66% relative to ReplaceMe on five models up to 27B parameters, and with a task objective it is up to 25 points more accurate given the same data. A second-order expansion of the loss yields a score that selects the repaired block from one forward and one backward pass of the intact model.

OpenAttention: One Percent Is Enough with Tail Proxies

Hyunwoo Yu, Jongbeom Lee, Hyunbin Kim, Sungjun Park, Jaemin Choi, Jincheol Yang — Fri, 23 Oct 2026 00:00:00 +0000

90% · Inference speed 90%, Compression 45%

Sparse attention reads only a small subset of the key-value cache, and how to treat the unread tail remains a key challenge. Discarding the tail biases the output, and sampling it as previous approaches do fails at reads of 5% and below, where the tail likely holds heavy hitters the selector missed. Our key observation is that the simplest possible proxy, one token that folds the entire tail into its count-aware mean, already yields lower output error than a sample at every read behind a practical selector. We further show in closed form that at low reads sampling nearly doubles the weight error of discarding while a single proxy stays below it. Building on these findings, we explore approximating the dense attention distribution with a few proxies that target the tail. We propose OpenAttention, a pipeline of clustering, ErrorGate and top-k selection followed by a count-aware mean proxy. To implement clustering and selection both efficiently and effectively, we design training-free Query-Geometry Clustering (QGC) and a Query-Geometry Selector (QGS), which compress the keys while respecting the query directions. Additionally, unlike prior methods that rank clusters by centroid alone, ErrorGate allocates the exact read by the estimated weight error of each cluster. On RULER-32K HARD with Llama-3.1-8B, OpenAttention retains 99% of dense accuracy at a 1% read. It also achieves state-of-the-art accuracy across models, context lengths and datasets. Our customized attention kernel, OpenAttention.flash, runs the decode attention stage 10$\times$ faster than dense attention at a 1% read.

BeNice! Closing Looping's Inference-Memory Gap via KV-Sharing

Hayden Prairie, Jacob M. Ryan, Taylor Berg-Kirkpatrick, Daniel Y. Fu — Fri, 23 Oct 2026 00:00:00 +0000

90% · Inference speed 90%

Looped architectures scale quality while keeping parameter count fixed by repeatedly passing activations through the same block of layers. During inference, each recurrence generates independent KV-caches, causing memory to grow linearly. To address this challenge, prior work has proposed KV-sharing (i.e., sharing KV-caches across depth), finding that, in some cases, the KV-cache of the final recurrence can be shared with the earlier recurrences without quality loss, a property we denote as KV-shareability. However, the mechanisms underlying KV-shareability remain poorly understood, with even identically trained looped architectures not consistently exhibiting it. In this paper, we study how KV-shareability can be enabled in looped models. We propose a two-stage method: in the first stage, we use stochastic depth during pretraining, which induces contractive residual dynamics that drive representations toward a common limit across recurrences. In the second stage, we propose BeNice!, a post-training method that broadcasts the final recurrence's KV-cache into the first recurrence, training it to produce compatible residuals while preserving the original model's quality. As a result, BeNice! enables KV-shareability while matching standard finetuning's unshared validation perplexity. Using BeNice!, we show that KV-sharing is complementary to alternative KV-compression methods, composing with memory-efficient sequence mixers (e.g., multi-latent attention, Kimi delta attention) and low-precision KV-cache quantization (e.g., fp8, mxfp8, nvfp4). Finally, in simulated low-memory, long-context inference, KV-sharing improves cache hit rates by up to 31.4%, increases throughput by up to 23.6%, and reduces p90 time-to-first-token by up to 83.6%.

★ To Keep or Not to Keep: Learning KV Cache Retention in Disaggregated LLM Serving Systems

Dong Liu, Yanxuan Yu, Eric Jiang, Shu Wang, Ying Nian Wu — Thu, 22 Oct 2026 00:00:00 +0000

96% · Inference speed 96%, Serving systems 48%

Disaggregated \textsc{LLM} serving separates prefill and decode across node pools, placing a network fabric between \textsc{KV}-cache computation and consumption. This breaks a key assumption of classical cache policies: a miss no longer incurs only local recomputation, but also network transfer to the decode node. Likewise, admitting a block to the global cache requires an additional transfer, so poor retention decisions waste both memory and bandwidth. We present \textbf{KVLearn}, a learning-based framework for cost-aware \textsc{KV}-cache retention in disaggregated serving. KVLearn includes: (i) a lightweight Prefix Reuse Predictor (\textsc{PRP}) that estimates reuse probability from structural and temporal prefix features without modifying model weights; (ii) a Cost-Aware Retention Score (\textsc{CARS}) that converts reuse probability into keep/admit decisions using recomputation, transfer, and storage costs; and (iii) an Adaptive Threshold Controller (\textsc{ATC}) that adjusts admission thresholds online from observed hit rates and memory pressure. We integrate KVLearn into a globally disaggregated serving topology and evaluate it on text and multimodal workloads, including large image/video-derived \textsc{KV} blocks with heterogeneous reuse patterns. KVLearn reduces end-to-end time-to-first-token by up to 56% versus No-Cache, 38% versus LRU-Pool, and 33% versus Mooncake-style baselines, while cutting inter-node \textsc{KV} transfer volume by up to 53% versus LRU-Pool. On MM-Session, throughput remains within ${\sim}5%$ of oracle. Code is available at https://github.com/FastLM/KVLearn.

★ Accelerating CPU Speculative Decoding via Unified Graph Execution and Transactional KV Isolation

Zhangdie Yuan — Thu, 22 Oct 2026 00:00:00 +0000

96% · Inference speed 96%, Serving systems 75%

Speculative decoding accelerates LLM inference, but deploying it on CPUs introduces severe systems bottlenecks: context-swapping latency between draft and target graphs, and memory allocation overhead during KV cache rollbacks. We introduce SubTernary, an execution framework that eliminates these costs through Unified Graph Caching and Transactional KV-Cache Isolation. By compiling both models into a single memory arena with a persistent thread pool and routing speculative tokens to a static scratchpad for zero-copy pointer resets, SubTernary avoids dynamic reallocation and thread synchronization entirely. We integrate a 256-bit AVX2 ternary SIMD kernel as a low-bit case study. Ablations reveal our systems runtime drives 35.0% of the throughput gains, while the ternary kernel adds 6.0%. SubTernary achieves a $1.35\times$ speedup over llama.cpp speculative decoding and $3.99\times$ over autoregressive decoding, maintaining $O(1)$ rollback latency.

★ Building a Nested LADDER for LLM Unstructured Post-Training Pruning

Thu, 22 Oct 2026 00:00:00 +0000

96% · Compression 96%, Inference speed 57%

Non-uniform allocation can improve post-training pruning of large language models, yet quality can still degrade sharply at high sparsity. To control degradation as sparsity increases, we introduce LADDER (Layer-Wise Allocation via Damage-Driven Evaluation and Refinement), which allocates sparsity incrementally. At each step, it evaluates pruning an additional fraction of weights from each eligible layer and commits the layer update that causes the least end-to-end degradation per removed parameter. This process produces non-uniform sparsity profiles and exactly nested masks across multiple budgets. We evaluate LADDER on five OPT, Llama, and Qwen models across 10-90% sparsity, using SparseGPT as its candidate-generation backend and comparing against uniform and adaptive allocation baselines. Results reveal a clear operating-regime trade-off: LADDER is comparable or slightly weaker at benign sparsity but becomes substantially more robust from 70% onward. This high-sparsity advantage largely persists after posterior distillation across four of five models and translates into improved retention of aggregate zero-shot accuracy.

★ Distilling Multi-Teacher Scoring Principles for Unsupervised Time Series Anomaly Detection

Dongchan Cho, Jiho Han, Juwon Hwang, Changhee Min, Minsang Kim, Namsoon Jung — Thu, 22 Oct 2026 00:00:00 +0000

95% · Compression 95%

Time series anomalies take diverse forms, and individual detectors encode different anomaly scoring principles that rarely cover all anomaly types. Recent unsupervised time-series anomaly detectors provide complementary scoring signals, yet deploying them as an ensemble preserves the inference and maintenance cost of every individual detector. We study whether heterogeneous teacher detectors can be distilled into a single lightweight detector in the unsupervised setting, where no anomaly labels are available to define a shared target. AnoMix addresses this by extracting weak pairwise supervision from the training-set anomaly scores of frozen teachers. It normalizes each teacher score sequence independently and trains the student with CoRank, a multi-teacher ranking-distillation objective that converts teacher score gaps into pairwise supervision through teacher consensus. Across seven benchmark datasets, AnoMix successfully integrates complementary teacher scoring principles into a single deployable model, improving over individual detectors and static score ensembles while using a compact student with under 80k parameters on the benchmark settings. A zero-shot variant further shows that the distilled ranking signal transfers to unseen datasets and can outperform large pretrained time-series foundation models for anomaly detection.

★ SPARE: Sparsity Allocation via Routing Entropy for Efficient MoEs

Yifan Yang, Hanling Yi, Dongquan Yang, Feng Lin, Rong Xiao, Aijun Li — Thu, 22 Oct 2026 00:00:00 +0000

95% · Compression 95%, Inference speed 84%

Mixture-of-Experts (MoE) architectures scale model capacity efficiently, yet their substantial memory footprint poses a major barrier to deployment. Expert pruning reduces this footprint, but most methods focus on which experts to remove and prune every layer equally. We propose SPARE (Sparsity Allocation via Routing Entropy), a training-free, selector-agnostic method that decides how many experts to remove from each layer using routing entropy, the entropy of tokens' normalized Top-K gate weights on calibration data. When a token's gate weight is spread evenly, removing one active expert discards little of it, so SPARE prunes layers with more diffuse routing more aggressively. A perturbation bound under random pruning formalizes this intuition, and controlled interventions with a fixed selector link concentrated routing to larger output changes. With the same selector and budget, SPARE improves average accuracy over uniform allocation on Mixtral, Qwen3, and DeepSeek models, especially at high pruning ratios. At 50% expert pruning, it raises the eight-task average of Qwen3-30B-A3B from 57.2% to 63.8%, and the math, code, and science average of DeepSeek-R1 from 80.1 to 82.8.

UlaCache: Unified Logical Addressing for Sparse LLM Serving

Isaac Zhu, Alan Cao, Hugh Yin, Rocky Wang, Orion Wang, Hscos Zhang — Thu, 22 Oct 2026 00:00:00 +0000

94% · Inference speed 94%, Serving systems 90%

Long contexts increase the memory cost of the key-value (KV) cache in large language model (LLM) inference. Sparse attention reduces attention computation and memory traffic by selecting historical tokens, but the retained KV cache must remain available for later decoding steps. We present UlaCache, a KV cache management system built around a unified logical address space spanning GPU KV pools and optional CPU memory. Stable logical references let the inference engine allocate and select KV independently of physical placement. Common allocation and sparse-access interfaces support data-parallel (DP) and tensor-parallel (TP) attention. For distributed storage, local GPU buffers retain selected KV across decoding steps, so only missing entries require transfers. When CPU storage is enabled, Parallel Bucketed LRU (PB-LRU) manages a GPU pool shared across requests. It updates recency metadata and selects eviction victims in parallel. UlaCache transfers KV and then updates the address mappings. With speculative decoding enabled under matched CPU and GPU memory budgets, UlaCache achieves peak decoding throughput speedups of 1.16 $\times$ (DP) and 1.70 $\times$ (TP) on H200, and 1.73 $\times$ (DP) and 2.51 $\times$ (TP) on B300, compared with the state-of-the-art baseline.

Shortcut-Resistant CAM Distillation for Long-Tailed Recognition

Wenhai Wan, Teng Zhang, Shao-Yuan Li, Xinrui Wang, Qiang-Sheng Hua, Songcan Chen — Thu, 22 Oct 2026 00:00:00 +0000

93% · Compression 93%

Real-world datasets often follow a long-tailed distribution, making generalization to tail classes difficult. We revisit this problem through the lens of shortcut learning, where models prefer the easiest predictive cues (e.g., background or textures) over object-centric semantics, especially under scarce and biased supervision. We find that this tendency is amplified for tail classes: limited examples often share similar contexts, making non-semantic signals highly correlated and thus tempting shortcuts, whereas head classes with diverse appearances and environments encourage more stable object-focused representations. Motivated by this observation, we propose Shortcut-Resistant CAM Distillation (SRCD), a plug-and-play framework that transfers object-focused explanations from head to tail classes. SRCD operates in the Class Activation Map (CAM) space, where a CAM provides a class-specific spatial evidence map for a prediction. SRCD aggregates CAMs from a small set of head-class candidates into a shortcut-resistant teacher using an energy-model weighting based on coherence and concentration, and distills it to the tail-class CAM. We provide a theoretical analysis that quantifies shortcut reliance as shortcut-region evidence mass in CAM space and shows that SRCD suppresses tail shortcuts. Extensive experiments on long-tailed benchmarks consistently improve strong baselines. The code is available at https://github.com/Haifeng3/SRCD.

How Far Can Foundation Models Be Distilled for Efficient Jet Tagging?

Jay Agarwal, Siddharth Khare, Dhruv Kumar — Thu, 22 Oct 2026 00:00:00 +0000

92% · Compression 92%

Jet tagging is one of the most important applications of machine learning at the Large Hadron Collider. Recent models build in more physics knowledge or rely on large-scale pretraining, but the experiments place strict limits on their inference cost. Knowledge distillation can transfer their performance to small students that are cheap to run. We ask which properties of a teacher help the student, how small a student can get and how close it can come to the state of the art. On a jet tagging benchmark, we distil six teachers that differ in architecture, symmetry and pretraining, including two foundation models, into five small students. Every teacher improves every student, and the gain depends mainly on the strength of the teacher. We then shrink our two strongest distilled students, Transformer and L-GATr-slim, in width and in depth. With 20 thousand parameters, a distilled L-GATr-slim outperforms every published jet tagger trained from scratch at a fraction of the inference cost of every teacher. Even with two thousand parameters, it matches ParT, a model with a thousand times as many parameters. We also study what the students learn from their teachers. Our benchmark of inference cost shows that students of the same size can differ by almost an order of magnitude in speed. We release the code and the weights of all trained models.

Stochastic Pruning of the DPO Computation Graph: 1.4x Faster Training and Lower Peak Memory

Yubing Wu, Wei Chen, Junmei Yang, Delu Zeng, Tiewen Pan — Thu, 22 Oct 2026 00:00:00 +0000

92% · Compression 92%, Inference speed 67%

Direct preference optimization (DPO) incurs training costs in output scoring, sequence execution, and backpropagation, so reducing supervision alone does not necessarily save computation. We introduce a unified framework for stochastic pruning of the DPO computation graph that independently samples response positions for scoring, response prefixes for execution, and preference branches for backpropagation. By making these selections before the corresponding operations, the framework omits computation at each stage of a training step. Our analysis characterizes the objective changes induced by position and sequence pruning and uses inverse-probability weighting to obtain a conditionally unbiased single-branch gradient estimator for the sampled objective. Controlled measurements show that position and sequence pruning primarily reduce peak memory, with most output-memory savings attributable to an implementation that retains full response supervision. Branch pruning yields a $1.426\times$ step-time speedup over a concatenated DPO baseline and reduces training time by about 30\% relative to stock DPO across two model families, with slightly lower mean accuracy and benchmark-dependent changes in open-ended scores. These findings relate the resource benefits of stochastic graph pruning to the computation omitted and the resulting changes in the training objective.

HAWK: Heterogeneity-Aware KV Cache Transfer for Efficient GPU-NPU Phase-Disaggregated LLM Inference

Thu, 22 Oct 2026 00:00:00 +0000

91% · Inference speed 91%, Serving systems 39%

Large language model (LLM) inference exhibits distinct computational characteristics across its two phases: prefill is compute-intensive, whereas decode is more memory-bound. As a result, AI accelerators achieve varying levels of efficiency across the two phases. GPUs are well suited to prefill, but are often underutilized during decode. By contrast, NPUs can utilize their compute resources more effectively during decode. Motivated by this observation, we explore a heterogeneous architecture that integrates GPUs and NPUs, mapping prefill to GPUs and decode to NPUs to achieve high performance and efficiency. However, realizing such heterogeneous integration is challenging because LLM phase disaggregation requires KV cache transfer from prefill to decode (that is, from GPUs to NPUs), which is complicated by differences in inter-device communication mechanisms and KV cache tensor layouts. To address this, we propose HAWK, a transfer mechanism that enables rapid KV cache transfer across heterogeneous devices. HAWK establishes a remote direct memory access (RDMA) path for direct cross-device transfers while handling inter-device tensor layout discrepancies as part of the transfer process. In particular, HAWK allocates tensor-layout-aware RDMA buffers on the NPU side to match the target tensor layout. Leveraging these buffers, HAWK orchestrates data transfer from a conventional contiguous GPU RDMA buffer to the corresponding NPU buffers, allowing transferred tensors to be consumed directly without additional data rearrangement while maintaining transfer efficiency. We implement and evaluate HAWK on a system built with commercially available GPUs and NPUs, demonstrating performance within 7.88% of a GPU-only system under the same SLO constraints while achieving up to 1.26$\times$ higher power efficiency and 1.72$\times$ higher cost efficiency.

★ Reconstruction-Optimized Expert Pruning for Mixture-of-Experts Language Models

Wed, 21 Oct 2026 00:00:00 +0000

97% · Compression 97%, Inference speed 80%

Mixture-of-Experts (MoE) models reduce computation through sparse expert activation, but deployment still requires storing a large number of expert parameters. We study post-training expert pruning using task-domain calibration data as a subset-selection problem: given a pretrained MoE layer, select a fixed number of experts that best reconstruct its original output. Exact subset search is combinatorial and becomes intractable for modern MoE models with many experts per layer. We develop OptEP, a scalable continuous relaxation that directly optimizes differentiable expert-selection variables, followed by a lightweight reconstruction-based update of the retained expert weights. Across multiple MoE architectures and tasks spanning mathematical reasoning and code generation, our approach improves over the strongest existing expert-pruning baselines, with absolute accuracy gains of up to 15.05% on Qwen3-30B-A3B, 14.55% on Qwen3-235B-A22B, 11.06% on GLM-4.5-Air, and 16.20% on Mixtral-8x7B. The improvements are particularly pronounced under aggressive pruning, demonstrating that reconstruction-driven expert selection and weight adaptation can preserve substantially more of the original model's performance while reducing the number of retained experts.

★ 3D-MoE: Complementarity-Guided Hardware-Software Co-Design for MoE Decode on 3D-Stacked NMP

Yudong Pan, Lian Liu, Shixin Zhao, Yintao He, Kaiyan Chang, Tianhua Han — Wed, 21 Oct 2026 00:00:00 +0000

96% · Inference speed 96%, Serving systems 80%

Mixture-of-Experts (MoE) have become a mainstream way to scale Large Language Models (LLMs), because each token activates only a small subset of experts, reducing inference cost. However, MoE decode remains highly memory-bound and exhibits strongly skewed, time-varying expert loads. 3D-stacked DRAM near-memory processing (NMP) provides the bandwidth that decode demands, yet its tiled organization makes load balancing expensive: expert parallelism (EP) preserves locality but suffers from hotspot-induced imbalance, while more distributed strategies weaken locality and increase NoC traffic. To address this locality-balance conflict, we observe that domain-specialized token routing creates offsetting expert-load fluctuations across decode steps, a property we call expert complementarity. Using this insight, we propose 3D-MoE, a hardware--software co-designed system for MoE decode on 3D-stacked NMP. 3D-MoE uses offline, one-time complementarity-guided grouping and locality-aware mapping to establish coarse-grained inter-group balance, handles residual intra-group hotspots through a lightweight Collaborative Memory-Network (CMN) that streams each hot expert's weight tiles once from its host DRAM channel to helper cores over a reconfigured local broadcast path, and uses cost-aware online remapping to remain robust to workload drift. Across MoE models, 3D-MoE achieves an average 4.63$\times$ MoE-layer speedup over EP and improves end-to-end throughput by 1.58$\times$--1.86$\times$ over the strongest NMP baselines, with 2.11\% additional logic-die area.

★ SliMOO: Interpretable Multi-Objective Evolutionary Search for LLM Depth Pruning

Guanchen Li, Yixing Xu, Xuanwu Yin, Dong Li, Emad Barsoum — Wed, 21 Oct 2026 00:00:00 +0000

96% · Compression 96%, Inference speed 93%

As large language models (LLMs) grow more powerful, their scale makes them costly to serve. Depth pruning offers a practical route to efficiency by directly reducing memory and latency without hardware-specific support. However, we observe that transformer layers exhibit markedly different task-dependent layer contributions, whereas existing depth-pruning methods ignore this heterogeneity and thus tend to optimize for a narrow domain or benchmark rather than preserve broad capability. This suggests that LLM depth pruning should be formulated as a multi-objective optimization problem rather than a single-score ranking problem. We propose SliMOO, an interpretable multi-objective evolutionary search framework for LLM depth pruning. SliMOO first uses single-layer removal probing to construct a task-layer importance map, which provides an interpretable view of task-shared and task-specific layers and serves as a search prior. It then performs a Pareto-aware evolutionary search for pruning masks, where candidates are generated under the guidance of the task-layer prior, scored by their deviation from the dense model on each task, and retained through NSGA-II environmental selection. Across the Llama-3.1 and Qwen3 model families, SliMOO consistently achieves stronger multi-task trade-offs than rule-based, greedy, and scalarized baselines. For example, on Llama-3.1-70B at 25% sparsity, it improves the math-domain average from 24.33% to 34.40% and the code-domain average from 35.03% to 39.14%. Code will be made available at https://github.com/AMD-AGI/SliMOO.

★ Beyond Policy Discrepancy: Capability-Aware Multi-Teacher On-Policy Distillation

Xuanfeng Yang, Le Gui, Jiaming Wang, Xuan Zhang — Wed, 21 Oct 2026 00:00:00 +0000

95% · Compression 95%

Multi-teacher on-policy distillation (MOPD) integrates domain specialists into a shared student, but determining how strongly each teacher should supervise remains challenging. Existing discrepancy-based allocation can assign strong supervision even when the student already matches or exceeds the teacher's task performance, since behavioral mismatch does not necessarily imply a remaining capability deficit. We propose Capability-Aware Multi-Teacher On-Policy Distillation (CA-MOPD), which combines specialist merging with capability-guided allocation to account for what the student has already learned. CA-MOPD first merges reinforcement-learning specialists into a strong student initialization, then uses verifier-estimated teacher--student capability gaps to adjust domain-level distillation strength. Supervision decreases as capability gaps close and increases when estimated student capability regresses, while dense token-level guidance remains provided by the teachers. Across six benchmarks spanning mathematics, code generation, and instruction following, CA-MOPD achieves an average score of 32.22, outperforming naive MOPD, Open-MOPD, and the merged initialization by 0.86, 0.83, and 0.74 points, respectively. Controlled comparisons further support the benefits of merged initialization and capability-gap guidance under matched training budgets.

MLIPS: Multi-Layer Indexer Pool Sharing

Samkit Jain, Sai Kishan Pampana, Shlok Agarwal — Wed, 21 Oct 2026 00:00:00 +0000

93% · Inference speed 93%, Serving systems 57%

Sparse attention limits the keys used by each query, but finding those keys can still require a full-prefix scan at every layer. We introduce \textbf{MLIPS (Multi-Layer Indexer Pool Sharing)}, a training-free method that shares candidate search while retaining each consumer's own scoring and selection. A producer publishes a top-$M$ pool larger than the attention budget $K$; subsequent layers select their own top-$K$ from that pool. Measurements across three sparse-attention model families show substantially higher coverage from broad pools than overlap between final selections. We implement layerwise sharing, persistent decode reuse, and a coarse producer in vLLM for DeepSeek-V4-Flash. Across four long-context suites, three consumer layers per producer give $1.14$--$1.40\times$ speedups and $14.8$--$27.4\%$ lower GPU energy, with observed accuracy changes of $-1.49$ to $+0.29$ percentage points in single runs. More extensive sharing exposes task-dependent quality losses. Sampled prefill traces also reveal recurring selected KV positions across query groups and layers. We characterize this shared working-set structure and use it to motivate Top-$G$, an exploratory bounded pool construction for prefill. Its deployment, task accuracy, and serving benefits remain open.

Weight Placement Is Not Enough: Synchronization in Multi-Node CPU Decoding of Ternary LLMs

Runguo Li, Dezhen Lu — Wed, 21 Oct 2026 00:00:00 +0000

93% · Inference speed 93%, Serving systems 51%

Ternary LLMs read only a few gigabytes per decoded token, yet on dual-socket servers with four NUMA nodes the fastest ternary runtime we found uses less than half of the memory bandwidth. Placing each weight on the node of the thread that reads it is necessary but not enough: in a decode engine with node-local weights that reuses ik_llama.cpp's kernels, threads on two generations of Xeon servers still spend 20-48% of each token waiting at synchronization points. Per-thread barrier traces show that most of the wait is for the slowest threads of each node; the barriers' own latency (microseconds each, but up to 12% of the token for the smallest model) and serial per-node work are smaller parts. On identical kernels, a Megatron-style column split beats a row-owned split without cross-node reductions in every configuration we measured, and isolating the number of barriers shows that most of its gain comes from its threads waiting less for each other and for their node leaders, and only a minority from having fewer barriers. Where decoding is DRAM-bound, the resulting engine is 1.5-2.1x faster than ik_llama.cpp with all cores and 1.5-1.9x faster than ik_llama.cpp at its fastest thread count (3.2x and 2.8x for a small model that runs from the L3 cache), at the same accuracy and with less energy per token.

Projected Distillation: On Learning to Solve Hard Problems from Close, Correct Targets

Wed, 21 Oct 2026 00:00:00 +0000

91% · Compression 91%

On-policy learning is central to post-training: training on a model's own generations limits distribution shift and helps preserve existing capabilities. On hard problems, however, on-policy RL provides no reward-based learning signal when all attempts fail. On-policy distillation provides next-token supervision from a teacher, but on hard problems, if a rollout goes wrong early, subsequent supervision is conditioned on prefixes that already contain the mistake. Correct behavior can instead come from a gold or teacher solution, but such off-policy targets can differ substantially from the model's generations and, as we show, may erode existing capabilities. We therefore treat on-policyness as a continuum and propose projected distillation, which minimally edits the student's failed attempts into correct targets. The method uses contrastive updates to favor corrections over failed attempts and refreshes these corrections as the student evolves, enabling near-on-policy learning even when all student attempts fail. Theoretically, we prove that the student's learning regret grows with the edit size: it scales linearly with the horizon under minimal correction, as in DAgger, and quadratically under full-answer replacement, as in behavior cloning. Empirically, across two students (Qwen3-8B and OLMo-2-7B) and five domains spanning math, code, science, and tool use, projected distillation gives the highest area under the specialization-retention frontier (best on 8 of 9 cells against GRPO and OPD, and 7 of 9 against SPOT and SODA). Our ablations highlight the importance of online corrections, the contrastive loss, and on-policyness: moving targets farther off-policy steadily worsens the specialization-retention trade-off, with retention falling from 0.916 to 0.744 when minimal corrections are replaced by student-agnostic gold solutions.

Beyond the Point Mass: Acceptance-Derived Weights and Distributional Losses for Speculative Decoding

Tianyu Wu, Zhuohan Wang, Jonathan Geuter, Yu Yao, Weifan Jiang, Chengxi Zhang — Wed, 21 Oct 2026 00:00:00 +0000

90% · Inference speed 90%

In speculative decoding, a lightweight draft model proposes tokens that the target model verifies in parallel, speeding up generation without changing the target distribution. D-PACE (Wu et al., 2026) trains such drafters with a loss that weights each position by its contribution to the expected accepted length. Both its weights and its cross-entropy (CE) loss use only the draft's probability of the recorded target token, treating the target as a point mass. Under rejection sampling, however, a sampled draft token is accepted with probability equal to the overlap of the draft and target distributions, one minus their total variation (TV). We derive D-PAL (Dynamic Position-wise Acceptance Loss) from the exact expected accepted length, using this overlap as the acceptance signal. This yields overlap-based position weights with log acceptance (LA) as the per-position loss, and the objective reduces to D-PACE for a point-mass target. Since the weights are detached, any per-position loss can replace LA in practice. We show that forward KL divergence and R\'enyi divergence of order 1/2 match or exceed LA, and that R\'enyi-1/2 provably bounds the true acceptance rate from above and below. Replacing CE with R\'enyi-1/2 improves acceptance length under both D-PACE's weights and ours. On six benchmarks with a Qwen3-4B target, D-PAL raises mean acceptance length over D-PACE from 4.50 to 4.57 at T=0 and from 4.20 to 4.28 at T=1, and consistently achieves higher throughput at both temperatures.

★ Factorization-Guided Training and Inference for Diffusion-Drafted Speculative Decoding

Jiakun Fan, Min Xiao, Qiang Wang, Neiwen Ling, Zhichao Liu, Zhihao Jia — Tue, 20 Oct 2026 00:00:00 +0000

97% · Inference speed 97%

Speculative decoding accelerates large language model inference through parallel verification of tokens proposed by a drafter. Diffusion drafters further parallelize proposal generation, but greater proposal capacity or higher mean accepted tokens (MAT) need not yield higher throughput. Because verification accepts only a consecutive prefix, early rejections can render later correct predictions ineffective, limiting MAT gains, while even higher MAT may not improve throughput if per-round latency increases. We introduce DScout, which uses two complementary factorizations to guide diffusion drafter training and inference for higher throughput. For training, DScout factorizes position-wise loss into inclusion, weighting, and prediction loss. This guides the design and combination of mixed-horizon training, intra-block unmasking, and acceptance-weighted loss to align supervision with prefix acceptance and improve MAT. For inference, DScout factorizes throughput as the ratio of MAT to per-round latency, guiding target-conditioning depth, drafter depth, and inference block size selection to identify a high-throughput configuration for each workload. Across six benchmarks, DScout achieves $3.84\times$ and $4.14\times$ mean speedups over autoregressive decoding on Qwen3-4B and Qwen3-8B, respectively, improving mean speedup over the state-of-the-art baseline by 9.1% and 7.1%. Across models and concurrency settings, our inference configuration policy provides up to 9.75% additional mean speedup over DScout's fixed configuration.

★ KVCompass: A Discovery Plane for Internet-Scale Prefix Cache Sharing in LLM Serving

Shiyi Ling, Fei Fang, Chen Qian — Tue, 20 Oct 2026 00:00:00 +0000

95% · Inference speed 95%

KV cache reuse is one of the most effective ways to reduce LLM inference cost, but today its benefits are largely confined to a single cluster. When a provider serves the same model across geographically distributed regions, each region may repeatedly prefill the same popular prefixes, wasting GPU time and increasing time-to-first-token (TTFT). Exploiting reuse across regions involves discovering which worker currently holds a reusable KV cache for each request's prefix. We show that this KV cache discovery problem is structurally harder than conventional object lookup. Unlike hash-placement mechanisms such as DHTs, where an object's location is a deterministic function of its key, KV caches are created on the workers that serve requests, and their holder sets change with subsequent requests and KV cache evictions. KV cache discovery therefore calls for decoupling KV cache identity from placement. We present KVCompass, a discovery plane for Internet-scale KV cache sharing. The discovery index supports variable-length prefix lookup with constant-time local table reads at the edge, allowing each request to be forwarded to the worker that best balances KV cache reuse against load and network distance. Evaluation includes a 64-worker geo-distributed simulation and a 14-worker three-continent real-machine testbed. Evaluation shows that KVCompass achieves $1.3-14\times$ lower P99 TTFT than the baselines, up to $3.7\times$ lower P99 end-to-end latency, up to $4.7\times$ the throughput, and a peak load factor within 2% of the Oracle floor.

★ LOOP: Recovery-Aligned Iterative Layer Pruning via On-Policy Distillation

Liang Xu, Shufan Shen, Zhongni Hou, Yufei Zhang, Wei Lin, Guojun Yin — Tue, 20 Oct 2026 00:00:00 +0000

95% · Compression 95%, Inference speed 81%

Layer-level pruning is attractive for large language model (LLM) deployment as it reduces the model size while preserving a dense architecture that can be efficiently executed on standard hardware. Current methods typically follow a one-shot paradigm, which first removes selected layers to reach a target model size and then performs additional training to recover model capabilities. However, pruning directly to the target size can substantially degrade reasoning capabilities, making subsequent recovery more difficult under a limited training budget. Moreover, this paradigm separates layer selection from recovery, limiting the use of recovery information to guide selection and potentially hindering the recovery of reasoning capabilities. To address these limitations, we propose LOOP, a Layer-level Recovery-Oriented On-Policy Pruning framework that interleaves pruning and recovery across multiple shots. In each shot, LOOP removes a single layer and performs short on-policy distillation, mitigating the abrupt degradation in reasoning capabilities associated with one-shot pruning. To guide layer selection in subsequent shots, LOOP reuses on-policy trajectories and optimization signals from recovery to jointly assess each remaining layer's reasoning contribution and recovery sensitivity, thereby making layer selection recovery-aware. Evaluations on several benchmarks show that LOOP consistently outperforms one-shot baselines in average accuracy across different model scales and pruning ratios.

Conjugate Compilation: Self-Improving Agentic Co-Compilation for Speculative Decoding

Tue, 20 Oct 2026 00:00:00 +0000

94% · Inference speed 94%, Serving systems 53%

As LLM adoption surges, scaling model serving requires unlocking stranded GPU capacity alongside deploying more GPUs. Speculative decoding brings two models into the serving process: a small draft model proposes tokens that a larger target model verifies. Yet compiling these models individually leaves their shared execution outside the compiler's optimization boundary. We introduce CONJUGATE COMPILATION, which reframes speculative decoding as a cross-model compilation problem and jointly optimizes the draft and target models within their shared execution rather than treating each in isolation. Our key insight is that cross-model compilation should be informed by the measured performance of the co-located shared execution it produces: an optimization that improves a kernel in isolation need not improve the models when they execute together. To expand cross-model opportunities, we devise operand layouts that interleave blocks from both models within common tiles, co-locating GEMM and non-GEMM work while preserving their mathematics. We realize CONJUGATE COMPILATION through a self-improving agentic search system in which an LLM edits cross-model optimization plans using the compiler and GPU execution as its environment. Each proposed plan is compiled and evaluated through its co-located shared execution, whose measured score, together with kernel-level verdicts and compiler feedback, informs subsequent proposals, while Monte Carlo tree search accumulates rewards to guide exploration. Across six single-GPU speculative-decoding configurations on NVIDIA B300 and RTX 5090 GPUs, CONJUGATE COMPILATION achieves a geometric-mean per-round speedup of 1.37x over conventional execution. Interleaved kernels raise this speedup to 1.45x, reaching 1.59x in the best case. CONJUGATE COMPILATION also improves GPU utilization on average by 1.46x on RTX 5090 and 1.50x on B300.

A Hybrid GPU and Processing-Near-Memory Architecture with Shared DRAM for Efficient Prefill-Decode Disaggregation

Sanghyeok Han, Inseong Choi, Byungkuk Yoon, Gyeonghwan Park, Jae-Joon Kim — Tue, 20 Oct 2026 00:00:00 +0000

94% · Inference speed 94%, Serving systems 50%

The growing adoption of large language models (LLMs) has significantly increased the compute and memory demands of inference serving. A key challenge arises from the mismatch between compute-bound prefill and memory-bound decoding: when executed together, the two phases interfere, while separating them via prefill-decode (PD) disaggregation leads to substantial resource underutilization. Specifically, prefill leaves memory bandwidth underutilized, whereas decoding underutilizes compute resources. Prior approaches, such as GPU Streaming Multiprocessor (SM) partitioning, attempt to improve utilization by spatially dividing GPU resources. However, these approaches remain fundamentally limited, as memory-bound decoding still requires a large number of SMs to saturate bandwidth, reducing the compute resources available for prefill and resulting in an inherent trade-off. To address this limitation, we propose DuoPD, a hybrid PD-disaggregation architecture that integrates lightweight Processing-Near-Memory (PNM) units into the HBM logic die. Unlike prior PIM/PNM designs that rely on costly internal bandwidth scaling and offload only partial decoding operations, DuoPD enables full decoding execution on PNM by leveraging bandwidth sharing with the GPU over a shared HBM interface. This allows compute-bound prefill and memory-bound decoding to execute concurrently with minimal interference, improving both compute and memory utilization. To realize this design, we develop a compact PNM microarchitecture, a TSV-level bandwidth sharing mechanism, and a hierarchical data mapping scheme tailored for hybrid GPU-PNM execution. Evaluations on server-scale LLMs demonstrate that DuoPD achieves up to 2.23x higher end-to-end throughput than conventional PD-disaggregation.

Parallel Language Models via Variational Decoding

Tue, 20 Oct 2026 00:00:00 +0000

94% · Inference speed 94%

Autoregressive language models emit one token per forward pass. This causes high inference latency, in particular as models are scaled up for quality. In this paper, we introduce Parallel Language Models (PLMs), which generate mutually consistent tokens in one forward pass by routing dependencies through latent auxiliary variables. PLMs are trained using a variational bound that reduces to an autoregressive loss plus a parallel decoding loss. Empirically, PLMs are on a faster and better Pareto frontier than autoregressive and discrete diffusion baselines. Pretrained PLMs achieve near-autoregressive pretraining loss on model scales of 80M to 3B parameters, and speedup follows a power law in model size. When post-trained, our 1.3B model matches or exceeds an identically trained autoregressive model on math, coding, and instruction-following benchmarks, speeding up inference $2.3\text{-}5.7\times$. Together, we show that parallel generation can be trained from a single training objective at autoregressive quality. We will make our code available at https://github.com/mandt-lab/plm.

CD-OPD: Contextual Digest for On-Policy Distillation

Mingyuan Li, Yucheng Li, Xiaxia Wang, Gong Cheng — Tue, 20 Oct 2026 00:00:00 +0000

92% · Compression 92%

Teacher takeover in on-policy distillation (OPD) can produce incorrect responses even when the student and teacher independently answer the same problem correctly, a failure we quantify using Collaboration Failure Rate (CFR). To address it, we propose Contextual Digest for On-Policy Distillation (CD-OPD). CD-OPD structures the earlier context of teacher guidance through a teacher-generated dependency graph, curating the required problem conditions and preceding reasoning as node summaries alongside the unchanged guidance in a Contextual Digest before student continuation. We further develop online meta-reweighting, which compares Digest segment gradients with a task gradient from independent student responses to improve student-only performance, favoring updates predicted to reduce task loss. Across eight benchmarks, CD-OPD improves average accuracy by 1.32--7.05 percentage points over representative existing OPD methods. It also reduces CFR by 27.4\%--56.4\% relative to existing takeover-based OPD methods.

Nash-MOPD: Optimizing Task Trade-offs in Multi-Teacher On-Policy Distillation

Siyuan Gan, Bo Yang, Linjian Meng, Boyan Wang, Xiaolong Qi, Jing Huo — Tue, 20 Oct 2026 00:00:00 +0000

92% · Compression 92%

Multi-teacher on-policy distillation (MOPD) aggregates the capabilities of multiple domain-expert teachers into a single student model. However, all domains in MOPD share the student's finite budget of optimization steps and rollouts. Fixed uses constant domain sampling ratios based on corpus sizes; by contrast, D$^3$-MOPD adjusts the ratios using the teacher-student reverse-KL divergence and its recent decrease. However, favoring large recent decreases can direct more training toward domains that already improve quickly, while domains that improve more slowly may receive too few examples. To address inflexible allocation and the risk of giving insufficient training to domains that improve more slowly, we propose Nash-MOPD, an adaptive domain scheduler based on a local minimax game using existing reverse-KL statistics. The equilibrium maximizes the smallest predicted proportional reduction under the local model and gives more prompts to domains that need more examples for comparable improvement. Consequently, Nash-MOPD responds to changing learning progress and supports domains that improve more slowly without extra teacher queries or rollouts. At step 100, Nash-MOPD exceeds both baselines' final mathematics and code checkpoint scores on DeepSeek-R1-Distill-Qwen-1.5B and reaches Fixed's on Qwen3-8B-Base. Furthermore, it leads eight of nine reported metrics and is best or tied best on all six mathematics and code accuracy metrics, respectively.

PeekKV: The First Content Token Already Knows Which Heads Matter

Minyoung Go, Yongjun Park — Tue, 20 Oct 2026 00:00:00 +0000

91% · Inference speed 91%, Compression 62%

The size of the KV cache grows linearly with context length, making it a major memory bottleneck in LLM serving. Motivated by recent progress in identifying the functional roles of attention heads, a growing body of work compresses the KV cache non-uniformly across heads. Existing methods either fix the important heads via offline profiling or select them using attention signals from the prefill stage. In both cases, however, accuracy fluctuates substantially across workloads. While investigating the cause of this instability, we find that the active heads shift between the prefill and decode stages even within a single input. Along the decode path, by contrast, the activehead composition remains stable across content tokens once format tokens are filtered out. Based on this observation, we propose PeekKV. PeekKV observes a single content token immediately after prefill - a one-token peek - to jointly obtain head-group scores and token-importance scores. Guided by these scores, PeekKV allocates KV cache budgets to head groups in proportion to their group scores and retains only the top-scoring tokens within each budget. PeekKV requires no offline profiling or training and thus applies directly to unseen workloads and new models. Across four public benchmarks, PeekKV achieves the highest accuracy in more than half of the per-model comparisons against state-of-the-art methods, and stays closest to the full-cache accuracy on RULER and LongBench. On Llama-3.1-8B-Instruct, PeekKV retains 95.7% of the full-cache accuracy at a 10% memory budget on RULER with 16K contexts and 95.9% at a 1.5% budget on LongBench.

The KV-Surgeon: Compensation-Aware Offline KV-Cache Compaction

Moritz Wagner, Christophe Roux, Max Zimmer, Sebastian Pokutta — Tue, 20 Oct 2026 00:00:00 +0000

91% · Inference speed 91%, Compression 42%

Offline compaction of the key--value cache (KV-cache) reduces the cost of repeatedly querying a fixed context by constructing a smaller cache once and reusing it subsequently. The central challenge of compaction is deciding which cache entries to delete and which to retain. Existing methods score each candidate for deletion on its own, with the rest of the cache being held fixed. However, softmax renormalization of the attention operation couples these decisions in a nonlinear manner: removing one entry changes the attention weights of every remaining one. In this work, we introduce KV-Surgeon, which uses calibration data to take interactions into account in a tractable way. We derive a quadratic surrogate that admits closed-form expressions for jointly making the optimal greedy removal decision as well as compensation within the remaining entries. On QuALITY, LongHealth, and Long Code Arena, and at every evaluated compaction ratio, KV-Surgeon substantially advances the cost--accuracy Pareto frontier of dense and mixture-of-experts models from 4B to 30B parameters. At $50\times$ compaction, KV-Surgeon improves question-answering accuracy by up to four points over attention matching (Zweiger et al., 2026), the state of the art, at comparable cost. Lastly, we introduce an even faster variant, which also exceeds the state of the art while requiring only two fifths of its cost.

Correctness-Dependent Asymmetric Target Distributions for Self-Distillation Policy Optimization

Suwan Yoon, Seongryong Jung, Changhee Lee — Tue, 20 Oct 2026 00:00:00 +0000

90% · Compression 90%

Self-distillation policy optimization provides dense token-level supervision for LLM post-training through feedback-conditioned self-teachers. However, directly matching the self-teacher can disrupt successful behaviors on correct rollouts while providing insufficient correction on incorrect ones. We propose Asymmetric Policy Optimization via Self-Distillation (APSD), which constructs correctness-dependent targets by mixing student and self-teacher distributions: interpolation preserves successful behaviors on correct rollouts, while extrapolation beyond the self-teacher strengthens correction on incorrect ones. We analyze APSD through divergence decomposition, reverse-KL barycentric geometry, and gradient analysis, highlighting the importance of gradient flow through student-dependent targets. Experiments under reinforcement learning with verifiable rewards and rich feedback show that APSD achieves strong overall performance against reinforcement-learning and self-distillation baselines and matches SDPO performance with 2.6x fewer training steps on average across Qwen3 model scales.

★ MemLLM: End-to-end In-Storage Acceleration of On-device Retrieval-Augmented Generation using Precomputed KV Caches

Seungkwan Kang, Hyungseok Ko, Heemin Kim, Myoungsoo Jung — Mon, 19 Oct 2026 00:00:00 +0000

96% · Inference speed 96%, Serving systems 43%

Retrieval-augmented generation (RAG) on edge devices enables personalized large language model (LLM) inference while preserving privacy. However, RAG's inherent length asymmetry makes the prefill phase the dominant bottleneck. While precomputing the KV caches offline alleviates the prefill cost on server systems, the multi-gigabyte caches over the narrow interconnect create a severe I/O wall for on-device systems. To mitigate this, we propose MemLLM, an end-to-end in-storage processing (ISP) architecture that executes the complete attention pipeline based on precomputed KV caches within the SSD. MemLLM utilizes per-channel accelerators to fully utilize the aggregated NAND channel bandwidth and accelerates all KV cache-related computations on the storage-side to minimize PCIe transfers. To efficiently handle the sparse updates required by RAG, MemLLM introduces a fused recomputation-query kernel paired with a hardware-based filter logic to halve the NAND read volume and eliminate write-back penalties. Furthermore, MemLLM's global SRAM-based stateful reduction engine enables a completely DRAM-less dataflow, thereby removing the internal memory bottlenecks and bounding the PCIe egress to an $O(1)$ footprint. A full-stack software stack allows asynchronous, zero-copy transfers from the ISP to the host and provides transparent execution within modern ML frameworks. Evaluated across eight real-world RAG datasets and six on-device LLM models, MemLLM achieves an average end-to-end speedup of 4.06$\times$ over the GPU baseline.

★ BS-Spec: BFP-Based Bit-Sharing for Self-Speculative LLM Decoding with LUT-Driven Mixed-Precision Acceleration

Jiwoo Kim, Gunho Park, Sunwoo Yoo, Baeseong Park, Dongsoo Lee, Youngjoo Lee — Mon, 19 Oct 2026 00:00:00 +0000

96% · Inference speed 96%, Serving systems 63%, Compression 33%

As large language models (LLMs) are increasingly deployed across diverse inference services, serving workloads span a wide range of batch sizes, context lengths, and performance requirements. Speculative decoding improves inference throughput, but conventional approaches require a separate draft model and KV cache, while existing self-speculative methods suffer from workload-insensitive draft configurations and inefficient mixed-precision execution. In this paper, we present BS-Spec, a block floating point (BFP)-based bit-sharing framework that jointly optimizes draft construction, workload-aware precision selection, and hardware execution. BS-Spec derives multiple draft representations directly from a single BFP target model by partially loading the most significant bits of weights and the KV cache while preserving shared exponents, eliminating separate draft-model storage. Offline calibration selects the weight and KV-cache precision pair that maximizes estimated speedup for each serving regime, and a precision lookup table applies the selected configuration at runtime. To support this adaptive execution, we design a mixed-precision BFP accelerator with a lookup table (LUT)-based variable-precision datapath for both draft and target computation. Experimental results show that BS-Spec achieves up to $1.9\times$ speedup over conventional autoregressive decoding and improves speedup over the evaluated bit-sharing self-speculative decoding baselines by $67.2\%$ on average, demonstrating the benefits of jointly co-designing bit sharing, precision adaptation, and specialized hardware for diverse LLM serving workloads.

★ A Stronger Teacher Within: Strengthening MoE Models via On-Policy Bidirectional Distillation

Haoyu Chen, Jun Bai, Yang Liu, Jianfei Zhang, Wenge Rong — Mon, 19 Oct 2026 00:00:00 +0000

95% · Compression 95%

Reinforcement Learning with Verifiable Rewards (RLVR) has substantially improved Mixture-of-Experts (MoE) LLMs, but its outcome-level rewards provide only sparse supervision. On-Policy Distillation offers dense token-level feedback, yet requires a stronger compatible teacher, while on-policy self-distillation relies on hand-designed privileged information. We observe that the MoE architecture itself offers a way to resolve this dilemma, as increasing the activation budget per token exposes latent expert capacity and yields a potentially stronger teacher within the model. However, naive expansion degrades performance because newly co-activated experts are not optimized to work together. We therefore propose MoEvolve, a bidirectional distillation framework following a recovery-then-improvement strategy: we first restore the activation-expanded high-$k$ model and further strengthen it by exploiting its additional active capacity, then reverse the direction of distillation to transfer the enhanced capabilities back into the original low-$k$ configuration, retaining its inference-time activation budget. Across diverse MoE models and benchmarks, MoEvolve improves the initial models by up to $22.2$ domain-average points and outperforms same-budget RLVR baselines by up to $6.5$ points. Notably, the final low-$k$ models often surpass their activation-expanded teachers.

★ Blitz: Accelerating the Diffusion Modules of AlphaFold3 and Boltz-2

Vinay Saji Mathew, Soundar Kumara, Gretta Kellogg, William K. M. Lai — Mon, 19 Oct 2026 00:00:00 +0000

95% · Compression 95%

Diffusion-based protein structure prediction models produce all-atom structures in about 200 denoising steps, and that cost has driven a number of few-step samplers. We present Blitz, a few-step distillation framework that uses a geometric noise schedule to distribute steps more evenly across noise scales, then learns those updates through multistep consistency training. Physical steering and target-free candidate selection help reduce the remaining geometric defects at inference. We apply Blitz to AlphaFold 3 and Boltz-2, reducing denoising from 200 steps to eight. Because published few-step results use different benchmarks and scoring procedures, we compare both distilled models with their teachers and other few-step methods under a common evaluation protocol covering structural accuracy, interface quality, ligand placement, and stereochemical validity. Blitz retains teacher-level structural accuracy, with AF3-Blitz outperforming all evaluated few-step baselines and exceeding its teacher on several accuracy metrics.

On the SPOT: Enhancing Speculative Decoding for Diffusion Language Models via Commit-Order Prediction

Eliezer Shlomi, Eliron Rahimi, Avi Mendelson, Chaim Baskin — Mon, 19 Oct 2026 00:00:00 +0000

94% · Inference speed 94%

Diffusion language models (DLMs) promise parallel text generation, but in practice face an unresolved speed-quality trade-off. Exact decoding fills in only one masked position per model pass, while faster multi-token decoding schemes are lossy. Speculative decoding is the standard lossless acceleration for autoregressive models. However, it does not transfer directly to DLMs, which must choose both what to generate and where. We show that this order is the missing piece: the model's own predictions from the previous pass already give the right token at 92-100% of positions, but not which position comes next. Based on this observation, we introduce Speculating Positions Order Trajectory (SPOT), which learns to predict this order. A 0.62M-parameter model reads only the target's output distributions and proposes a tree of candidate continuations, and the target verifies all of them in a single pass by checking whether each equals its own next state. On the verification side, our analysis shows that prior block-diffusion verification approaches either assume $T=0$ or a sequential proposal interface. To support tree-based speculation at any temperature, we propose using coupled-noise match verification for diffusion blocks, following shared-Gumbel-noise coupling approaches. Our comprehensive experiments show that SPOT cuts target passes by up to 6.8x and reaches up to 3.2x the throughput of standard decoding, beating the calibrated and training-free drafting baselines in every task and budget at $T=0$ and $T=1$. In addition, we show that the learned drafter transfers across models, demonstrating successful transfer across six different DLMs. Project page: https://eliezer318.github.io/SPOT-dLLM/

DIFFLET: A Fast Inference Engine for Diffusion Transformer Models on AWS Trainium

Bin Ma, Min Yu, Vijay Kodamalla, Jialin Liu, Dong Li — Mon, 19 Oct 2026 00:00:00 +0000

93% · Serving systems 93%

On AWS Trainium2, every component of a diffusion model runs as a program compiled ahead of time for fixed shapes, while pipelines differ across models, prompt lengths across requests, and caching decisions across denoising steps. We present Difflet, a serving engine in which five model pipelines share one compilation backend, joint attention is split across cores by attention head, and a host controller decides at each denoising step whether to run the denoiser or reuse its earlier output. A small probe program computes the caching signal on the device and returns one number per step, and a resident model process keeps every program loaded across requests. Across five image and video model families, fused attention accelerates denoiser steps by 1.2-4.3x over the same framework using PyTorch/XLA attention, and at 32.8k visual tokens it is the only attention that compiles. On four of the five models, resident serving is 1.5-3.4x faster than standalone runs that reload weights from a warm file cache for each generation.

GR-OPD: Group-Relative Sequence Credit for On-Policy Distillation

Yushu Wang, Renkai Yang, Jingcao Xu, Zili Geng, Haihong Tang, Yaohui Jin — Mon, 19 Oct 2026 00:00:00 +0000

92% · Compression 92%

On-policy distillation (OPD) provides token-level teacher feedback, yet common sampled-token updates assign credit locally and omit the effects of earlier actions on subsequent feedback. We introduce \method{} (Group-Relative On-Policy Distillation), which decouples feedback computation from credit assignment by aggregating token feedback into calibrated sequence-level credit. We establish a theoretical basis for this approach by showing that, under ideal on-policy assumptions, broadcasting raw summed feedback to every response token yields an unbiased estimate of the negative gradient of the sequence-level reverse-KL objective. To support effective learning in practice, \method{} discounts later feedback that may become less reliable under distribution shift and uses same-prompt rollout groups to calibrate teacher guidance. It preserves relative response preferences while retaining a controlled fraction of the group mean, which we term collective teacher pressure. Across five Qwen3 teacher--student settings and six mathematical reasoning benchmarks, \method{} outperforms vanilla OPD in macro Avg@K in every setting, including gains of 10.9 and 11.5 percentage points in two settings with matched thinking patterns. It also maintains stable training under mismatched thinking patterns, where vanilla OPD collapses. Ablations show that retaining an appropriate amount of collective teacher pressure can improve learning, while excessive collective teacher pressure or an unrestricted feedback horizon can destabilize training.

Next Token or Vector? Learning How Far to Decode

Mon, 19 Oct 2026 00:00:00 +0000

91% · Inference speed 91%

Standard autoregressive decoding advances one token at a time, even when several upcoming tokens can be predicted together. We propose CADENCE (Context-Adaptive DEcoding), a generation method that learns both what to draft and how far to draft. Gaussian prediction heads represent future spans as continuous latent vectors, and a shared decoder converts the selected vector into token predictions in parallel. A learned router selects the prediction length from the model's own hidden states and predicted Gaussian parameters, adapting generation granularity to the current context with single-token generation as a fallback. The backbone verifies each proposed continuation and retains the prefix that matches its greedy predictions. Experiments with Llama-3.1-8B-Instruct on Alpaca, GSM8K, and HumanEval show that router-selected positions yield higher draft acceptance than applying the corresponding prediction length everywhere. CADENCE achieves up to 1.3x higher generation throughput than standard greedy decoding.