2026-09-29

LLM-based agents execute long-horizon tasks through repeated model calls interleaved with tool execution and user interaction. As each call extends the history accumulated in previous turns, prefix caching avoids repeated prefill of the agent's entire context. However, the aggregate cache footprint grows with context length and concurrency, forcing serving systems to reclaim cached KV tensors. We identify recency synchronization: accesses to an agent's cached history refresh its cache nodes together, causing least recently used (LRU) eviction to concentrate on a few agents' private histories. When a fully evicted agent returns, it must recompute nearly its entire accumulated context, producing a large time-to-first-token (TTFT) outlier even if most other requests retain substantial cache reuse. We present RR-EVICT, a fine-grained prefix-cache eviction strategy that distributes reclamation across idle agent trajectories. RR-EVICT visits agents in round-robin order and evicts a tail chunk from each, preserving reusable prefixes for more agents when capacity remains for idle state. Returning agents can reuse these partial histories and recompute smaller missing suffixes. The policy requires no prediction of future arrivals or tool latency. We implement RR-EVICT in SGLang and evaluate conversational and coding-agent workloads under colocated and prefill-decode-disaggregated serving. Compared with LRU, RR-EVICT reduces P99 TTFT by up to 75.4% and P99 uncached prompt tokens by up to 65.7%.

Large language model (LLM) serving is increasingly constrained by the GPU memory consumed by key-value (KV) caches. Existing compression, eviction, and offloading techniques alleviate this pressure, but serving runtimes typically treat only the configured target KV representation as execution-ready. Under memory pressure, this target-only contract can turn KV shortage into request stalls and preemptions. We present ElasticKV, a mixed-fidelity KV runtime built on the observation that target fidelity need not gate execution. ElasticKV introduces a compact intermediate KV state, making fidelity a runtime-managed execution property. To realize this state in a paged serving runtime, ElasticKV combines (i) a pair-structured layout that turns fidelity reduction into reusable GPU capacity, (ii) a dual-mode attention backend that directly consumes the compact state while preserving the native target-only path, and (iii) pressure-aware fidelity management that adapts KV fidelity to memory pressure. Our extensive evaluation across diverse workloads, model families and scales, and GPU platforms demonstrates the effectiveness and generality of ElasticKV. Under high concurrency, ElasticKV achieves 3.8-4.0$\times$ lower time-to-first-token (TTFT) and 9.1$\times$ lower P90 TTFT than vLLM while preserving generation quality.

Multi-LoRA agent systems enable efficient role specialization by sharing a common backbone model. However, each agent repeatedly processes the growing shared trajectory and constructs its own KV cache, introducing substantial memory and computation redundancy in long-horizon tasks. Existing KV cache sharing methods reduce this repeated prefill, but they either require additional training or architectural constraints or retain substantial model computation. Moreover, direct cache reuse causes the current agent to rely on cache states generated by the previous agent's adapter, weakening the role-specific behavior encoded by its own LoRA. We present PReCache, a training-free KV cache sharing framework with two designs, namely PreLRShared and ReBaseShared, that share the base cache computed using the pretrained weights and precompute a compact agent-specific low-rank (LR) cache. To remove repeated prefill, PreLRShared precomputes each agent's LR cache when the shared context is first processed, allowing the current agent to use its own LR cache without reprocessing context processed by previous agents. To improve sharing accuracy, ReBaseShared reconstructs the shared base cache from adapter-free hidden states, reducing the remaining error caused by the previous agent's adapted representation. To minimize its reconstruction cost, we propose two inference schemes tailored to single-stream inference and concurrent serving, performing the same reconstruction after each agent's turn or alongside its execution, respectively. Across multiple models and agent benchmarks, PreLRShared achieves up to a 3.1x TTFT speedup and a 2.3x improvement in per-request throughput over inference without KV cache sharing. ReBaseShared best preserves accuracy overall among the evaluated cache-sharing methods, with an average drop of only 1.1 points relative to inference without cache sharing.

Prompt-specialized multi-agent systems enable multiple agents to share a model while performing complementary roles to solve complex tasks. However, agent-specific prefixes change the KV cache generated for the same shared context, causing each agent to repeatedly prefill the growing context and construct a separate cache with high computation and memory overhead. Selective recomputation reduces this redundancy but still retains substantial model execution, while existing delta correction methods either support only recurring context relations or maintain memory-intensive online correction states for dynamically changing context. For first seen shared context, these methods also construct a reference cache outside the agent workflow, and an approximate correction at the first agent affects the outputs passed to subsequent agents. We present KVCMAS, an online KV cache correction framework that represents cross-agent cache deviations using compact low-rank states and seamlessly chains corrections along the agent workflow without an additional reference prefill. This design supports dynamically changing shared context while preserving an exact first-agent cache. Across multiple language and vision-language workloads, KVCMAS matches or improves the accuracy of prior KV cache sharing methods while achieving the lowest TTFT under highly concurrent serving. Under controlled serving traces, it provides a 2.0x TTFT speedup over inference without KV cache sharing and reduces peak GPU memory by up to 3.7x relative to a prior KV cache correction method. These results establish KVCMAS as an accurate and scalable KV cache sharing approach for prompt-specialized multi-agent serving.

Mixture-of-experts (MoE) language models often exceed the memory of a single GPU. Expert offloading keeps most experts in host memory and loads them on demand, so decoding speed depends on how many experts each token must fetch. Caching and prefetching reduce this cost only as far as the routing allows. Router-only fine-tuning can reshape the routing to reuse experts, but it keeps the experts frozen, so they cannot adapt to the tokens the new routing sends them. We propose MaskCoFT, a masked co-adaptive fine-tuning method that trains routers and experts together with the cross-entropy loss alone. During fine-tuning, a learnable binary mask restricts the Top-K routing of each layer to a subset of experts, and the experts adapt to the tokens redirected to them. At inference, the learned mask becomes a soft prior that re-ranks experts, so every expert remains selectable. We simulate a GPU cache of 4 experts per layer for Mixtral-8x7B and 12 for DeepSeek-V2-Lite. MaskCoFT cuts expert fetches per token by 23.7% and 10.1% relative to the base model. In real offloading system serving, it lowers the time per output token by up to 16.4% and 5.5%, respectively. Its average accuracy over nine benchmarks stays above the base model by 0.92 and 0.53 points.

Existing LLM serving systems virtualize and optimize KV-cache memory, but treat model-weight memory as fixed throughout execution. Recent work on multi-precision model representations challenges this design by allowing a single stored model to support both full-accuracy and lower-precision execution, making the effective weight footprint runtime-dependent. This creates an opportunity under bursty workloads, where temporary spikes in KV-cache demand often determine throughput and SLO compliance. We present DPS, a dual-precision LLM serving system that turns weight memory into an elastic resource: under normal load, DPS serves the full-accuracy model; under KV pressure, it switches to a nested, lower-precision variant and repurposes unused weight memory for KV cache blocks. DPS is built on Semi-Unified Memory (SUM), which partitions the weight region into a persistent lower-precision sub-region and a shared region that alternates between residual weight tensors and KV-cache blocks, preserving compatibility with paged KV-cache management. We implement DPS on top of vLLM and evaluate it across both dense and MoE models and various production workload traces. Our results show that \sysname improves sustained throughput by $2.1$--$3.3\times$ and effective pass@1 by up to $+41$\,pp over Static FP16, while preserving FP16-class accuracy.

Pipeline parallelism can improve prefill throughput by processing multiple request chunks concurrently across different stages of the model. However, keeping the pipeline fully utilized requires efficient scheduling and request preparation. In systems where stages retain and evict cache state independently, a local cache hit does not guarantee that the same prefix can be reused across the pipeline. Here, coordination overhead can impede request admission cadence and thus reduce overall throughput. In this paper, we present WavePP, a prefill runtime built on top of TensorRT-LLM that addresses these challenges by overlapping request admission with pipeline execution. WavePP asynchronously finds a prefix that can be reused across all stages, protects the cached state, and reserves space for the remaining input while earlier requests continue to execute. It subsequently plans the chunk sizes of each request dynamically to maximize pipeline fill. Each stage then completes the local preparation before executing the request. In the same system and pipeline topology, WavePP improves TensorRT-LLM's prefill throughput in 37 of 40 tested settings on GLM 5.2 and MiniMax M2.7. At concurrency 128 with high cache reuse, these changes increase throughput by factors of 2.91 and 2.02, respectively. Across 28 Kimi K3 settings, WavePP also has the highest measured throughput in all 18 settings at concurrency eight or higher, compared with tensor/expert-parallel and pipeline-parallel baselines from TRT-LLM, SGLang, and vLLM.

Attention often concentrates on a small subset of tokens in the context, but which subset matters changes from one query to the next. To exploit this changing structure, we introduce SANTA++, a training-free stochastic attention method that uses representative keys for memory-efficient selection without scanning the entire key-value (KV) cache. Cached keys are organized into teams, and the query scores one representative from each team to decide which teams to sample. We compute exact attention scores within the sampled teams and reweight each team's contribution by the inverse of its inclusion probability. This importance sampling correction estimates attention over the full cache, with a sampling budget that lets us trade memory reads for accuracy. Remarkably, with 32 or 64 sampled teams, SANTA++ uses 16% to 22% of dense attention's KV reads and retains 94% to 99% of the dense-attention baseline's scores on LongBench v2 and HELMET's retrieval-augmented generation subset, and 85% to 91% on RULER, with Qwen2.5-7B-Instruct at 32K context. With 31 sampled teams, our GPU implementation delivers a $1.69\times$ attention speedup over the dense FlashAttention baseline at 32K context. By reducing the number of cache entries read, SANTA++ in principle complements architectures with compressed KV representations, such as multi-head latent attention. Our kernels are available at: https://github.com/OPUSLab/santapp-kernel-demo.git.

RAG and retrieval-based agent memory both inject retrieved content into LLM prompts, as document chunks and recalled memory records, respectively. The same content can recur across requests at different prompt positions or after different preceding contexts, preventing reuse through conventional prefix caching. Our characterization finds that records recurring outside the matching prefix account for over 70% of injected memory tokens in agent-memory workloads. Composable KV-reuse methods enable reuse in such cases, but online serving introduces a management problem: a recurring unit's KV states may not yet exist, may have been evicted, or may reside on another node. We present Tessera, a disaggregated serving system that makes retrieval the control plane for KV reuse. By exposing the context units needed before model execution, retrieval allows Tessera to combine current demand with retrieval history, KV residency, and generation load to coordinate cache management and request routing. Generation nodes concurrently prepare locally cached, remotely cached, and missing states, while retaining newly computed states off the request's critical path. Across RAG and agent-memory workloads, Tessera lowers mean TTFT by up to 3.6x over SGLang and LMCache with EPIC at matched request rates, and sustains low TTFT at rates where the baselines saturate, while matching the answer quality of the underlying composition policy.

On-policy distillation (OPD) trains a student on its own rollouts using dense supervision from a teacher. In multi-turn environments, a mistake at a critical decision step can redirect the subsequent rollout toward poor outcomes. We use low teacher confidence on student actions to select high-uncertainty steps for correction. In a controlled ALFWorld study, a single teacher correction at a low-confidence step improves subsequent student behavior and task success, motivating selective intervention during distillation. We propose UOPD, an uncertainty-aware intervention method for on-policy distillation. At low-uncertainty turns, UOPD executes student actions and applies the standard OPD loss. At high-uncertainty turns, it samples and executes teacher actions and trains the student to imitate them through supervised fine-tuning, which minimizes forward Kullback-Leibler divergence in expectation. UOPD utilizes adaptive uncertainty thresholds to target a scheduled intervention rate. Empirically, we evaluate UOPD across a broad range of agentic tasks, including ALFWorld, WebShop, and Search, demonstrating its superior performance over OPD methods and their variants. UOPD improves WebShop score by up to $15.8\%$ relative to standard OPD.

Reusable prefix key-value (KV) caches can outgrow GPU memory in large language model (LLM) serving. A memory-semantic flash hierarchy offers SSD-backed capacity with a limited fast tier, but a logical KV hit is not necessarily ready for GPU retrieval. Demand staging exposes SSD latency, whereas immediate staging can reserve fast-tier capacity long before retrieval begins. We present TempoKV, a timing-aware resource-commitment layer that separates early knowledge of reuse from the acquisition of staging resources. It records reusable-KV hits as metadata-only claims and requests commitment when the runtime-estimated time until retrieval falls to the storage-estimated time needed to make KV resident and protected against eviction. These estimates adapt to runtime progress and staging state, while commitment remains subject to available protected capacity. We implement TempoKV in vLLM and LMCache on an SSD-backed CXL memory device without changing request scheduling. Across two models and three prefix cache ratios, TempoKV reduces protected fast-tier byte-time per request by 63-91% versus immediate staging while retaining much of the serving benefit of advance staging. In a fast-tier capacity sweep, output throughput and p95 time to first token (TTFT) remain nearly unchanged as capacity decreases from 100 to 25 GiB. Compared with unmodified LMCache's Device-DAX L1 configuration, TempoKV reduces p95 TTFT by up to 48.0% and increases output throughput by up to 27.8%.

Reinforcement learning can turn one language model into several specialists, each excellent at a single skill such as mathematics, coding or following instructions, but users need one model with all of these skills. Multi-teacher on-policy distillation (MOPD) merges them by letting the specialists teach one student: the student answers each prompt, and the specialist for that prompt's domain gives feedback on every token. This routing decides which specialist teaches, but not how strongly its feedback moves the shared student. In Qwen3.5 models at three sizes, we find that MOPD's student does not beat one taught by the best single specialist and gains little of the mathematics specialist's advantage. The feedback is unbalanced: instruction-following feedback is several times more spread out than mathematics feedback and dominates the student's updates. We propose Domain-Normalized MOPD (DN-MOPD), which keeps the routing and rescales each domain's feedback by its measured spread. On six public benchmarks, DN-MOPD improves the average score over MOPD at every size, across three random seeds and under two answer-length limits, and recovers most of the lost mathematics gain. Controls with fixed domain weights show that the gain comes mainly from turning down instruction-following feedback rather than turning up mathematics alone, and that fixed weights close to those DN-MOPD measures perform comparably. Combining specialists therefore requires deciding not only which one teaches, but also how strongly its feedback counts.

Verbal feedback can identify errors and prescribe corrections, providing rich supervision for language-model post-training even when reliable programmatic verifiers are unavailable. Such feedback, often generated by a capable model, can be used to condition the teacher in on-policy distillation, which trains the student to match the teacher's predictions on student-generated rollouts. However, this approach can transfer teacher preferences that the feedback did not motivate, while leaving much of the feedback's guidance unused. We find that both problems come from the standard on-policy distillation objective, specifically the divergence it minimizes and the distribution it uses as its target. Our proposed method addresses both limitations. First, to isolate the information conveyed by the feedback from the teacher's inherent preferences, we treat verbal feedback as evidence for or against the hypothesis that a particular token comes next at a given prefix. We then adopt a probabilistic confirmation framework which uniquely determines an ordering over the vocabulary based on the teacher's predictions before and after it receives feedback. Using a confirmation score consistent with this ordering, we construct a target distribution within a trust region of the student. Second, to learn from guidance that student rollouts can leave unused, we derive a simple shared-rollout estimator of a symmetric divergence between the student and target distributions over rollouts, reusing student and feedback-conditioned teacher rollouts in both directions through importance weighting. Empirical evaluations show that our method outperforms the common on-policy distillation recipe and a recent contrastive variant on knowledge-based and agentic benchmarks.

On-policy distillation (OPD) trains a student language model with dense feedback from a stronger teacher on student-generated trajectories. Yet standard OPD weights token-level distillation terms uniformly, implicitly treating local teacher preference as a proxy for correction utility. A decision's task value, however, depends on how the student completes the subsequent reasoning. This mismatch can cause imitation to suppress viable student strategies or reinforce paths the student cannot reliably execute. Verified trajectory outcomes provide complementary evidence about continuation quality, but do not directly identify the utility of individual decisions. We introduce Reward-Aligned Reweighting for On-Policy Distillation (R$^{2}$-OPD), which uses outcome agreement and the magnitude of teacher--student disagreement to continuously reallocate teacher supervision. It gives reward-aligned corrections greater relative influence while retaining dense feedback, moving beyond uniform imitation and hard filtering. Our analysis formalizes the mismatch between local teacher preference and student continuation value and establishes sufficient conditions for reallocation to improve first-order task progress over uniform OPD. Across seven mathematical reasoning benchmarks, R$^{2}$-OPD achieves the highest average accuracy among the compared training methods in both cross-size and same-size distillation. It outperforms standard OPD on all seven benchmarks, with average gains of 3.5 and 2.4 percentage points for 1.7B and 4B students, respectively. An extension to code generation yields an average gain of 1.6 percentage points over standard OPD. These results highlight outcome-guided supervision allocation as an effective way to translate dense teacher feedback into stronger student performance across model scales and task domains.

Amortized neural likelihoods enable computationally expensive inference for models with analytically intractable or unspecified likelihoods, but their black-box nature limits interpretability. We introduce a symbolic distillation pipeline that converts trained neural likelihoods into explicit, interpretable expressions optimized for efficient parameter estimation. Our approach uses a recovery-directed objective to guide symbolic regression toward expressions that preserve parameter-recovery accuracy rather than merely approximating the likelihood function. Candidate expressions are evaluated on held-out datasets and selected using a criterion that jointly accounts for expression complexity, parameter-recovery performance, and distributional distance from the learned likelihood. We evaluate the pipeline on the diffusion decision model, a classical cognitive model, whose analytically tractable likelihood provides ground truth for controlled evaluation. The proposed recovery-directed objective improves parameter recovery over standard symbolic-regression objectives. The resulting symbolic likelihoods enable over 100 times faster parameter evaluation than both neural likelihoods and, when available, the exact likelihood, while maintaining a manageable loss in precision. We further demonstrate these computational benefits in Bayesian hierarchical inference on empirical data. Our pipeline provides a lightweight interface for integrating symbolic distillation with existing neural-likelihood estimation methods and can be adapted to a range of simulation-based inference settings.

Unified multimodal models combine understanding, generation, and editing within a single network, offering a promising foundation for versatile multimodal applications. However, growing multimodal contexts make KV cache storage and access increasingly costly. Existing KV cache compression methods are typically tailored to specific tasks and single-modality caches, while overlooking changes in cache importance across tasks and timesteps. However, in unified multimodal models, each task involves multiple KV cache types, and both their composition and dynamics differ across tasks. As a result, a single compression policy overlooks task- and type-specific requirements, leading to the loss of critical information and degraded quality across tasks. Based on these findings, we propose UniCache, a training-free framework for task- and type-aware KV cache compression. UniCache identifies the cache segments activated by each task and assigns suitable compression policies through offline calibration. It coordinates their parallel execution under a shared storage budget through attention-guided allocation and task-aware temporal scheduling. Experiments show that UniCache achieves $5\times$ KV cache compression for understanding and editing and $2.5\times$ for generation with negligible quality loss, while increasing throughput by up to $1.78\times$ in long-context settings, significantly improving the practicality of scaling unified multimodal models to longer context.

Between them, the members of a research group or a circle of friends own several consumer computers, none large enough to run a capable large language model. Existing systems pool such capacity across open swarms anyone may join, which a group admitting only trusted machines cannot use. Bounding membership removes what they depend on: a swarm holds each part of the model on several peers and routes around a slow one. A bounded session must use every device it admits. Its pipeline advances at the pace of whichever device received a share it cannot serve quickly, so the division has to be right before serving begins. We propose Kafila, whose protocol assembles a ring from behind NATs, preferring direct paths and relaying where traversal fails, while its planner measures each device's memory bandwidth, capacity and reachability, divides the model exactly for a fixed ring order, and places the head, which holds the embedding and output projection, together with that division rather than beforehand. On machines with different capabilities across three fleets, from a shared LAN to five devices spanning two continents, Kafila shortens the slowest pipeline stage by up to $5.2\times$ against the even split of pipeline parallelism, as in GPipe, and up to $3\times$ against the memory-proportional split of personal-device inference, as in exo, keeps 75 to 87 per cent of the committed hardware doing work where those divisions fall below half, and serves a model no uniform split can place on the fleet at all. What that is worth to a user depends on how much of a token is computation rather than network. Where the members share a network the same division returns $1.56\times$ the throughput of a uniform split and $1.25\times$ of a memory-proportional one, and under four concurrent users that lead compounds to $3.2\times$ rather than fading, each user served at almost the rate of one.

On-policy distillation (OPD) trains a student on its own trajectories with dense teacher supervision. Recent work on OPD for multi-turn autonomous agents often treats large teacher-student token-level distributional gaps as promising intervention points, linking larger gaps to a greater need for correction. Yet, our empirical analysis reveals a supervision-benefit mismatch: large gaps can be benign, while small gaps can be outcome-critical. Teacher-student gaps capture differences at the current turn, whereas the benefit of teacher guidance depends on how the current student interacts with the environment afterward. The student may still succeed despite choosing an action that differs from the teacher's, while a teacher-preferred action may lead to a state from which the student cannot complete the task. Local gaps alone are therefore not enough to determine whether teacher guidance benefits the current student. Effective supervision should instead emphasize guidance that the current student can translate into better final task outcomes. Accordingly, we propose Outcome-Guided On-Policy Distillation (OG-OPD), which applies trajectory-relative weighting to teacher supervision and calibrates these weights using final task outcomes from paired student continuations. This calibration selectively strengthens supervision on the student's original trajectories at turns where teacher guidance benefits the current student. Across ALFWorld, ScienceWorld, and WebShop, OG-OPD consistently outperforms baselines under diverse settings. It improves task success rates by 3.6-17.7 percentage points over vanilla OPD and by up to 7.0 percentage points over the strongest baseline.

Leech-lattice quantization gives good quality at two bits per weight, but its codebooks hold more than 10^14 points, too many for a lookup table. Our earlier kernel expanded the codes at load time and read 4.804 bits per weight from GPU memory for 2 bits of code. We present Tetra, a new codebook on the same lattice. A 24-weight block still takes 48 bits, most of which index a 64-state trellis of the Golay code and one shared 16 KiB table. The kernel decodes a block with six table loads and two small lookups inside the matrix-vector product, and reads 2.148 bits per weight. For full models, we retrain one scale per matrix row, store the matrices that lose the most as 4-bit integers, and pay for them with 4-bit embedding tables. Our Qwen3-4B, 8B and 14B files hold 2.73, 2.70 and 2.73 bits per parameter over the whole model. They score 63.37, 69.58 and 75.66 on the full MMLU test set, 4.76, 4.21 and 2.46 points below 4-bit AWQ at 5.3 to 6.0 bits per parameter. They generate 113.8, 95.0 and 57.2 tokens per second in our engine. On GSM8K, through the served kernel, they lose 9.63, 4.62 and 3.26 points to FP16. At 4B our file scores 23.6 points above llama.cpp's IQ2_XXS (2.48 bits per parameter). Every number we measured for a table or figure comes from one NVIDIA L40S GPU. We preregistered the main experiments.

We study on-policy distillation (OPD) through the lens of reinforcement learning, establishing a connection between the reverse-KL objective in OPD and KL-regularized policy optimization. Building on this connection, we introduce Least-Square Policy Distillation (LSPD), an RL-inspired framework that brings optimistic exploration and off-policy data reuse from value-based RL into policy distillation. LSPD preserves policy diversity through exploration while improving rollout efficiency by repeatedly learning from previously collected trajectories. Our theoretical analysis connects LSPD to optimistic value-based learning and shows that its idealized formulation achieves a sharp $\tilde{\mathcal O}(\log K)$ regret bound under online exploration. Empirically, LSPD consistently outperforms existing distillation baselines across six mathematical reasoning benchmarks and diverse teacher-student settings, with average gains of +1.59 points in Avg@16. Remarkably, through Pass@k evaluations up to k=64, we found that LSPD better preserves policy diversity by achieving stronger performance as k grows. Its fully off-policy variant achieves comparable performance to vanilla OPD using only the first 25% of rollout batches. Together, these results provide an RL perspective on OPD that offers both a principled interpretation and a practical route toward more effective and rollout-efficient language model distillation.

Hybrid transformer architectures that replace most softmax attention layers with linear attention offer transformer-level quality at a fraction of the memory cost. Rather than pretraining such models, a growing body of work distills them from already trained full-attention transformers. However, these distilled models often collapse on long-context retrieval and reasoning tasks, particularly when operating in thinking mode, where the efficiency gains of hybrid architectures matter most. Since linear attention layers must compress context into a fixed-size state, their errors compound over long sequences. As off-policy distillation never teaches the student model to recover from this drift, tasks that necessitate longer sequence lengths become especially challenging. We introduce On-Policy Attention Linearization (OPAL) in which the hybrid attention student samples its own long-context trajectories and receives dense supervision from the frozen full-attention teacher. Applying OPAL to Qwen3-4B and MiMo-7B-RL-0530, we recover $87$--$94\%$ of full-attention performance on commonsense reasoning, $100\%$ on needle-in-a-haystack (NIAH) retrieval, and $83$--$93\%$ on mathematical reasoning with only 3B training tokens. We achieve these results without supervised fine-tuning (SFT) or reinforcement learning with verifiable rewards (RLVR). Compared with the strongest prior linearization method, which recovers $68\%$ of its teacher's retrieval performance and $21.6\%$ absolute average mathematical reasoning accuracy, OPAL fully recovers retrieval and achieves $67.6$--$72.2\%$ on math reasoning.

On NVIDIA Blackwell B200, tensor-core throughput outpaces special-function exponential throughput by more than two orders of magnitude, exposing exponential evaluation in fused attention kernels. A pretrained Transformer, however, may not need it evaluated accurately at every element. We characterize what a pretrained model does need by approximating softmax at inference in ten frozen decoder-only models (0.5B-72B). The number of positions the softmax map assigns probability to and within-row resolution can be cut substantially, yet uniform weighting of the same positions is damaging. Where a fixed resolution budget is placed matters as much as its size, with resolution near the row maximum consistently favored. Perturbations matched on scalar distortion produce model-dependent responses of opposite sign. These findings motivate Rowmax-PoT, a coarse logarithmic weight representation anchored at each row maximum, and Rowmax-H15, its hardware specialization in FlashAttention-4. On B200, the patched FP8 attention forward is 12.4% faster at causal 8K and 25.8% faster at non-causal 8K in host-side call-latency measurements; board energy per forward falls by 8.4% at causal 16K. Measured separately on the BF16 kernel path at 2K, Rowmax-H15 increases perplexity by 0.091-0.492% across five models from three families.

LLM agents resend their whole conversation on every turn, and most of it was already processed on the previous turn. Serving systems avoid recomputing it by caching its key-value (KV) state and, when GPU memory runs out, by offloading that state to host memory. For agents, offloading gives inconsistent results: on the same coding-agent workload it speeds up one deployment, slows down another, and changes nothing on a third, even where loading a token back is several times cheaper than recomputing it. The reason is that cached state must survive until it is used again. While one agent waits for its tool, the server processes the contexts of all other agents, so an agent's prefix is reused only if the host tier holds the reusable context of the whole agent pool, which we call the reuse working set. A smaller tier keeps writing state that is evicted before anyone reads it. We present EfficientAgent, which sizes and manages the host tier by this working set. A stack-distance model estimates the working set from agent histories to size the host tier; its predictions, made before the experiments, located the capacity at which offloading starts to pay. When the tier is too small, a runtime policy stops writing large refills of evicted context and keeps extending prefixes that are still cached; when the tier is large enough, it writes everything. On SWE-bench Verified coding agents, a host tier sized to the estimated working set cuts recomputed prompt tokens by 93% and end-to-end time by 39%. With a small fixed tier, the policy cuts recomputation by 35%; with a large tier, it avoids the 4.3-fold increase caused by always filtering writes. Across three GPU types and two models, offloading pays off when the GPU has little compute per byte of host bandwidth and the host tier holds the working set. Code is available at https://github.com/KunmingSHAO/efficientagent_release.

Multi-teacher on-policy distillation (MOPD) combines independently developed domain teachers into a single student by distilling their predictions on student-generated samples. We study a setting where teachers share a reference model but undergo different post-training procedures, and find that MOPD can struggle to recover some teacher capabilities. Because distillation occurs on student-generated prefixes, the student initialization can strongly affect subsequent recovery. However, initial benchmark performance is not a reliable predictor of a good MOPD initialization. For example, merge initialization can start below SFT warm-up yet finish higher after MOPD. We further find that effective merging depends on both the relative teacher contributions and the overall merge scale, with some strong configurations lying outside the simplex of convex parameter averaging. Thus, selecting a good merge initialization requires evaluating not only its immediate performance but also the learning it enables under MOPD, making one-shot coefficient search difficult. We propose Iterative Merging for MOPD (IM-MOPD), which starts from a uniform merge and progressively adds task-vector increments for under-recovered domains during distillation. In a 5-domain setting, IM-MOPD achieves higher average normalized recovery than MOPD with either uniform merge initialization or SFT warm-up, showing that effective teacher contributions can be determined progressively during training.

Most post-training quantization pipelines fit each weight matrix to its pretrained counterpart, one matrix at a time. Whether that proxy tracks what an attention block actually computes, or how errors in the Q, K and V projections compound inside the softmax, is rarely checked. We write the objective on the attention output instead, over all three projections at once, and reuse it throughout the pipeline. JAB defines one scalar loss over the joint Q, K, V weights of a block, evaluated against the block's real causally-masked attention output, and uses it twice: to fit the quantized weights (GPTQ warm start, then STE with learnable scales), and to score the block for a multiple-choice knapsack allocation. On attention-only quantization of Mistral-7B this works. At 3 bits JAB recovers 77-90% of the gap between uniform GPTQ and full precision, and its sensitivity estimate tracks an oracle costing 73 forward passes to within a fraction of a point. It stops working once MLP layers enter the allocation. A role-aware offset rule needing no sensitivity estimate at all beats JAB on GPT-2's MLP and on the full Mistral-7B model: with a 3-bit floor it quantizes 96.4% of the weights to 4.5 bits per parameter at 6.933 perplexity, within 4.4% of full precision (6.643) at 3.56x compression, against 7.158 for JAB at the same budget. Which matrix a weight sits in matters more than any sensitivity estimate we computed. Two things came out sideways. Block-local reconstruction is an unreliable proxy for end-to-end perplexity: one run improved a block's own objective 4.6x while perplexity rose 32x, which is why every allocation here is validated end-to-end. And on attention-only quantization, fine-tuning moved weights farther from their pretrained values while pulling attention outputs closer, with net gains. Post-training seems to recover attention behavior, not weights.

At each decoding step a language model attends over the key-value (KV) cache of every earlier token, so at long context the attention call is bounded by memory bandwidth. Sparse attention reads only a subset of keys chosen by a cheap score estimate, and most methods give the unread tokens zero weight. The output then draws on only a small fraction of the KV cache, and accuracy drops at small budgets, most on tasks that aggregate information across the context. An inverted-file product-quantization (IVF-PQ) index over the cached keys computes an approximate score for every indexed token in order to rank them; after ranking, those scores approximate the attention logits of the tokens left out. PQ-HSA (hybrid sparse-approximate attention) attends the selected tokens with their original keys and values, and the unselected tokens, the background, enter the same softmax through those scores, summed per inverted list and multiplied by the list's mean value. At 128K and a 1-2% retrieval budget, PQ-HSA is more accurate than Quest and SnapKV on Llama-3.1-8B and Qwen3-30B-A3B and stays close to full attention in macro accuracy; with the same selector, the background term raises macro accuracy on the 8B model from 0.71 to 0.83. In the same 128K setting, inside vLLM on one NVIDIA H20, the decode attention call runs 1.6x faster than the FlashAttention-3 kernel; the speedup grows with context length, and a cost model fitted on 8B to 30B models gives the context length at which it begins. A vLLM plugin runs PQ-HSA on two engine versions without changes to the engine source; code is available at https://github.com/KunmingSHAO/pqhsa_release.

Pretrained autoregressive Transformers represent a large sunk investment in compute. Existing conversion methods reuse that investment by changing either the architecture (attention to recurrence) or the objective (next-token prediction to denoising), never both. We convert Qwen3 teachers at 1.7B and 8B into attention-free, bidirectional, gated-delta-rule diffusion students in three stages, so that each capability can be traced to the stage that kept or lost it. Language modeling transfers only partially and in-distribution; in-context retrieval does not transfer. On a multi-query recall probe where the teachers score 0.34-0.58, both converted students score 0.000, and diffusion pretraining alone does not restore retrieval. A retrieval curriculum in the final stage, which gradually lengthens the gap between a key-value table and the queries that address it, restores it only stochastically: on a fixed schedule, one seed in three learns to retrieve. Advancing the gap only while a running accuracy estimate stays above a threshold works for all three of those seeds, holds on real text, and carries unchanged to 8B, where two of three seeds succeed. The third had not learned within its fixed 16k-step budget: retrieval switches on abruptly at a seed-dependent step (6.5k and 11k in the other two), so a fixed budget can cut a late run off. One boundary survives every intervention: every model that learns retrieval scores 0.000 on tokens that never appeared in a retrieval episode, and an arm that resamples the key and value tokens every batch shows this is a coverage limit, not memorization of particular bindings. Separately, we convert a 7B code model into a 3:1 recurrent-attention block-diffusion hybrid over 85k steps and report two negative training results.

AlltoAllv communication is a critical primitive in distributed large-model inference, particularly for mixture-of-experts (MoE) models. The growing adoption of PCIe GPU systems for cost-efficient inference makes AlltoAllv performance on these systems increasingly important. Without a dedicated scale-up interconnect (e.g., NVLink or Infinity Fabric), PCIe GPU systems carry both intra-node and inter-node traffic through the PCIe hierarchy, where concurrent transfers can contend for PCIe link bandwidth. This link contention, compounded by skewed traffic distributions and dynamic traffic demand, makes efficient AlltoAllv scheduling challenging. Existing approaches are either poorly suited to PCIe GPU systems or incur substantial schedule synthesis overhead that reduces their practicality in real-world deployments. We present VarioPath, an efficient AlltoAllv scheduling framework for PCIe GPU systems. It combines an offline topology-aware analyzer with an online demand-aware scheduler. The analyzer records contention-free transfer patterns as AlltoAllv channels and exploits topology symmetry to build a compact catalog for efficient search. The online scheduler decomposes each AlltoAllv invocation's demand across a sequence of channels, adapting to rapidly changing and skewed traffic while incurring low planning overhead. Evaluation on four platforms (up to 256 GPUs) shows average AlltoAllv speedups of 5.88x over FAST and 1.72x over DeepEP. End-to-end experiments show that VarioPath reduces Qwen3 inference latency by up to 27.2% and Wan2.1 generation latency by 6.1%.