2026-09-30

Agentic sessions driven by Large language models (LLMs) often alternate between model inference and tool use, accumulating long histories across successive rounds. Serving these sessions efficiently requires reducing attention computation and retaining history key-value (KV) caches to avoid recomputation. Recently, frontier open-source LLMs adopt sparse attention to reduce computation by selecting only part of the history, while SSDs provide a cheaper alternative to CPU DRAM for storing KV caches. However, sparse KV selection depends on the ad hoc intermediate values during model inference, so it forces SSD reads to lie on the inference critical path. These reads are further slowed by fragmented accesses and read-write interference in SSDs. To address these challenges, we present Janus, an agentic serving framework for sparse attention LLMs with SSD-centric KV storage. Janus focuses on append prefill, which processes each round's newly added inputs and accounts for most history KV loading. To move SSD reads out of the critical path, Janus runs the model's own KV selection module on earlier intermediate values, predicting KV demand without additional training. The predicted reads overlap with model computation, and any prediction misses are fetched before attention executes to preserve model outputs. To improve SSD efficiency, Janus coalesces adjacent reads, packs scattered KV pages into sequential writes on the CPU, and limits background writes while reads are active. Across three models and three agentic traces, Janus outperforms existing works by up to 1.57-3.69 times (1.22-1.85 times on average) in terms of the time to first token latency, while maintaining decode efficiency.

Dynamic layer skipping reduces LLM computation by allowing each token to execute only a subset of the model's layers. However, existing skippers rely on specialized generation loops and do not integrate with modern serving engines. As a result, fewer executed layers do not necessarily translate into lower serving latency: FlexiDepth skips 8 of Llama-3-8B's 32 layers on average, yet its standard generation loop decodes 14.6--21.0% more slowly than the base model. We present vSkipper, a virtualization layer that makes dynamic layer skippers pluggable in serving engines while preserving continuous batching, fixed-shape batches, paged KV caching, and captured decode graphs. At each routed layer, vSkipper groups tokens by the skipper's decision and uses routed execution only when predicted to be profitable. We implement vSkipper in SGLang and evaluate the released FlexiDepth checkpoint against upstream SGLang under identical prompts, arrivals, output lengths, and launch settings. At the knee of upstream's load curve, vSkipper reduces mean end-to-end latency by 36.8% on GSM8K and 13.6% on BBH. Under saturation, it increases request throughput by 11.3% and 7.4%. Serving adds no statistically resolved quality loss beyond the checkpoint's own. Across synthetic skip policies, two Qwen3 skippers, and three GPUs, we demonstrate reuse without workload-specific tuning. To our knowledge, vSkipper is the first system to realize serving-efficiency gains from per-token interior layer skipping within a modern LLM serving engine. The code is open-sourced as an SGLang fork at https://github.com/AKafakA/sglang-vskipper/tree/vskipper-ref

No single way of parallelizing attention serves large language models well under all loads. Low concurrency favors tensor parallelism, many independent requests favor data-parallel attention, and long prompts favor context parallelism. Reasoning, agentic, and RL-rollout workloads make a fixed choice untenable: a batch that begins as many short requests ends as a few very long ones, so the best layout changes while the same requests run. Serving engines nevertheless fix one layout at launch, because changing it has meant draining requests and restarting workers. We present SPLASH, a serving system that switches the parallel layout of attention while requests are running. It builds on one observation: modern attention, with few or no KV heads, decouples where a request's KV cache lives from how attention weights are sharded. This has two consequences. First, layouts differ only in who owns the weights and the cache, and most of that state already sits where the next layout needs it; SPLASH reuses it, moves the rest in the background of ongoing inference, and hands off at a batch boundary, making a switch nearly free: its median overhead is under 0.51% of the step it runs in. Second, the decoupling exposes a layout that existing engines lack: Decoupled Ownership Parallelism (DOP) shards attention weights as tensor parallelism does while keeping each request's cache on a single owner as data-parallel attention does. DOP replicates neither, offers 27-60% more KV capacity than data-parallel attention, and gives the scheduler a choice when KV memory limits admission. A transition-aware scheduler follows the best of the four layouts as load changes. On B200 GPUs serving GLM-5.3, SPLASH improves end-to-end serving throughput by 1.3-1.73x over fixed-layout deployments, and the same layout regimes appear with DeepSeek-V3.2 on H200 and GLM-5.3-Flash on DCU.

LLMs trained with Chain-of-thought excel in reasoning capability, but often come with excessive token cost. We find that the core of reasoning capacity lies in the Thinking model's weight component within the null space of a projection defined by the corresponding Non-thinking model's dominant singular directions, and removing the subspace component can largely improve reasoning efficiency without hurting the accuracy gained during thinking-mode post-training. Unlike existing efforts that mostly operate within the dominant subspace, we are the first to unveil the critical role of the null space and harness it for model optimization. Motivated by this finding, we propose Spectral Null-Space Swap ($S^3$), a training-free composition of paired Non-thinking and Thinking checkpoints. Our method keeps the Non-thinking model inside its own dominant subspace and takes the Thinking checkpoint outside it, improving reasoning efficiency while maintaining accuracy. We extensively evaluate $S^3$ on 2B-30B dense and mixture-of-experts (MoE) architectures spanning 28 evaluation environments across mathematical, multimodal, and audio reasoning domains. $S^3$ establishes new empirical Pareto Frontiers among training-free model composition strategies: across all settings, it reduces inference token overhead by an average of 27.4% compared to full Thinking models while simultaneously improving overall task accuracy by 1.0 percentage point (e.g., yielding +8.3% accuracy on HMMT25 alongside a 33.0% token speedup). We further use attention entropy for explanation and find that the retained component produces more concentrated attention, and we use a simplified analytical model about optimization to demonstrate why null-space can effectively reduce attention entropy, thereby improving the efficiency of reasoning.

Uniform random rotations are an effective preprocessing step for quantization: they make normalized coordinate distributions approximately Gaussian, enabling the use of codebooks optimized offline. We introduce TORQUE, a framework that improves on previous quantization works that use random rotations by jointly optimizing how many and which coordinates to preserve at high precision both before and after rotation, under a fixed overall expected bit budget. Intuitively, before rotation, preserving large input coordinates at high precision can reduce overall error by preventing the rotation from spreading their values across many coordinates. Likewise, after rotation, preserving a small fraction of the largest-magnitude coordinates at high precision allows the remaining values to be quantized more accurately using codebooks optimized offline for the resulting truncated Gaussian distribution. We derive a quantization error upper bound and prove that top-$k$ pre-rotation retention minimizes it for each $k$. This reduces the search over coordinate subsets to an optimization over $k$, enabling a fast optimizer that uses offline codebooks and parallel parameter selection for practical implementation. We demonstrate an improved tradeoff between reconstruction accuracy and storage cost through numerical evaluation under the Gaussian model and experiments on nearest-neighbor retrieval, KV-cache compression, and activation compression.

Large language models are increasingly expensive to serve. In large-scale serving systems, autoregressive decoding is often bottlenecked by transferring model weights from accelerator high-bandwidth memory into on-chip SRAM. Mixture-of-experts (MoE) models reduce computation by activating only a small subset of experts per token, but this sparsity does not translate directly to batched decoding. Different requests select different experts; therefore, the combined active set across many concurrent requests can span a substantial fraction of the expert pool and require significantly more expert weights to be transferred. Most expert-reduction techniques make retention decisions independently for each token and therefore do not address this batch-level expansion. More recently, batch-aware methods have attempted to coordinate expert use across concurrent requests and reuse experts already fetched for the batch. Yet their selection criteria are based primarily on router rankings or expert statistics collected during calibration. Consequently, these criteria are not directly tied to the output error caused by dropping an expert, nor do they capture how an expert's contribution changes across tokens at inference time. We instead rank experts according to how much their removal would change the MoE-layer output. To apply this criterion during serving, we train a lightweight linear predictor during calibration that estimates the expert removal cost for each incoming token, and develop custom GPU kernels for cost prediction and expert selection. Across three MoE architectures, BASE improves the quality-efficiency tradeoff without retraining. On Qwen3-30B-A3B, it improves average accuracy by 29.5 points over the strongest baseline at comparable throughput under a tight expert budget. At a higher expert budget, it is 60% faster than dense inference while remaining within 0.4 accuracy points.

On-policy distillation (OPD) trains compact language agents with teacher feedback on student-generated trajectories. In multi-turn tasks, compounding errors can move students beyond the teacher's effective supervision. We introduce Graph-Conditioned On-Policy Agent Distillation (GC-OPD), which enriches an off-the-shelf teacher's scoring context with execution evidence. A graph indexes repeated teacher executions by shared states while preserving complete successful and failed histories. After each student episode, GC-OPD retrieves current-state references or historical alternatives and combines them with student hindsight to score the original thought-action tokens. Using the same original teachers, GC-OPD improves mean success over vanilla OPD from 24.70% to 48.78% on ScienceWorld (4B student), from 53.36% to 85.26% on ALFWorld Unseen, and from 29.10% to 37.65% on WebShop. At matched student sizes, it also achieves higher mean success than every evaluated OPD baseline using GRPO-trained teachers on ScienceWorld and ALFWorld; the strongest such ScienceWorld 4B baseline reaches 46.66%. GC-OPD requires no task-specific teacher optimization.

Parallel speculative drafting generates multiple candidates in one backbone pass, but independent token selection can produce inconsistent continuations that shorten the accepted prefix. Existing methods mostly leave conditional decoding to a lightweight module after the backbone, which limits the flow of predecessor information to successors. Our analysis of DFlash shows that early positions already form recoverable predictions in shallow layers, and that accurate adjacent predecessors help successors more when they enter earlier. We therefore propose DSpine, a drafter with causal conditioning injection throughout the backbone: at every layer, gated adjacent injection writes each predecessor's predicted feature into its successor, so the causal conditioning chain unfolds over network depth while all positions update in parallel. A unified transfer space built from the target model's output embeddings unifies layer-wise injection with predecessor-conditioned decoding, and layer-wise output-embedding supervision promotes the formation of predicted features in shallow layers. Fused kernels and a transition cache execute both efficiently in parallel within SGLang. Across seven math, code, and chat benchmarks, DSpine achieves the longest acceptance length at both temperatures on Qwen3-4B and Qwen3-8B. At temperature zero on Qwen3-8B, it raises the seven-benchmark mean from DFlash's 3.77 to 4.82 (+27.8%); in SGLang serving tests, it delivers 23.3% higher throughput than DFlash on average.

Text-to-image users often provide concise and underspecified prompts, whereas generative models benefit from detailed textual conditions for reliable instruction following. Existing systems bridge this gap with Prompt Enhancers (PEs) that rewrite raw prompts at inference time, introducing additional latency and leaving prompt elaboration external to the generator. We instead view enhanced prompts as privileged training information and ask whether their benefits can be internalized. We propose Prompt-Enhanced On-Policy Self-Distillation (PE-OPSD) for text-to-image flow-matching models. During training, a raw-prompt student follows its own generation trajectory, while an enhanced-prompt teacher provides vector-field targets at the states visited by the student. This on-policy supervision distills the behavior induced by enhanced prompts into the raw-prompt student without requiring additional text--image pairs. At inference, both the PE and teacher are removed, and the student generates directly from raw prompts. Across multiple model families, PEs, and benchmarks, PE-OPSD achieves the strongest aggregate prompt fidelity among the evaluated baselines, yields positive aggregate visual appeal gains, and retains the base-model inference efficiency.

Self-distillation turns knowledge distillation into a closed learning loop and offers a path toward recursive self-improvement. Without an external teacher, however, the model must determine both what information can improve its supervision and which induced changes should be learned. Existing methods typically improve teacher-generated data or select training examples in isolation, leaving the information transferred between these stages unmeasured. We introduce InFlow, a retrieval-guided on-policy self-distillation framework that models this process as potential-to-realized information flow. InFlow first retrieves potentially informative sources using certainty-calibrated hidden-state trajectories, then measures their realized effect through the Jensen--Shannon divergence between the teacher's initial and retrieval-conditioned answer beliefs. Examples with larger belief shifts are selected for on-policy distillation. Our analysis formalizes the information optimized by retrieval and selection and relates the answer-level shift to the teacher--student distillation gap. Across four open-weight language models and three knowledge domains, InFlow achieves the strongest cross-model average among the compared selection methods, with ablations supporting both stages of the framework. Our code is available at https://github.com/1240148048/INFLOW.

Recent work on anchored diffusion language models improves denoising by shaping an intermediate latent space with supervised important-token targets. In this work, we introduce time-based (self-supervised) anchoring, which learns and reuses latent anchors without requiring such targets. Our key observation is that anchors encode persistent properties of the clean sequence, such as its semantic intent, global structure, or intermediate plan. Although their hidden representations become stale as the token canvas evolves, their semantic content remains useful across nearby diffusion times. This is implemented through a two-stage architecture consisting of a relatively expensive anchor network that generates the latent cache state and a lightweight denoising network that intelligently combines the cached latent state with the current state at each reverse step using a fusion module. This gives anchoring a latent-space caching interpretation: the anchor network is evaluated periodically, while its cached representation is reused across multiple reverse steps. We instantiate this framework as TADM:Post-train, which time-anchorizes pretrained DLMs, and TADM:Pretraining, which learns time-based anchors during pretraining. Applied to DiffusionGemma-26B, TADM:Post-train improves throughput by approximately 49% to 79% on several math, code, and STEM benchmarks (GSM8K, AIME26, GPQA-Diamond, LiveCodeBench-v6, HumanEval, MMLU-Pro). TADM:Pretraining reduces Transformer-layer computation by up to 38% relative to a standard single-stage DLM, achieves up to 73% higher measured throughput than ADLM.

Industry applications often demand low-latency classification, yet current large language model (LLM) approaches remain poorly suited for latency-critical applications. Existing prompting and constrained decoding produce verbose, multi-token outputs that require expensive token-by-token generation, while encoder-based models achieve faster inference but sacrifice task flexibility. We propose Koa-action, a framework for low-latency atomic actions -- fast, single-step decisions such as classification, semantic endpointing, Boolean checks, and scoring -- formulated as constrained generation with single-token outputs. By introducing atomic label tokens and applying supervised fine-tuning, our method reduces classification to a deterministic one-step decoding problem. Across standard benchmarks, Koa-action delivers competitive accuracy with consistently low and stable latency. On a production intent-routing benchmark, Koa-action reaches 85.5% accuracy -- competitive with the strongest frontier models (Claude-4.8-Opus, Gemini-Pro-3.1) and ahead of GPT-5 and Gemini-2.5-Pro -- while answering in about half a second, several-fold faster than every frontier model (up to ~7.5x at the median) under identical serving conditions. Against the dedicated single-token system Jev/TypeSafe, Koa-action is competitive on accuracy and faster at the median, while also handling multimodal inputs and multi-label outputs that single-label text systems do not.

On-policy distillation (OPD) trains multi-turn language agents with dense teacher supervision on student-generated responses. However, standard think-then-act rollouts require lengthy reasoning before each short action, delaying environment transitions and experience collection. Generating actions directly reduces this delay but can degrade rollout quality. To address this, we propose ActFirst-OPD, an act-first, reason-later training framework that decouples environment interaction from full-response generation. The student infers and executes actions through reference-conditioned inverse dynamics using its current interaction context and a reference next observation, and switches to autonomous next-action prediction when the resulting transition deviates from the reference trajectory. From the collected interaction contexts, the student asynchronously generates full think-then-act responses for token-level teacher supervision. Experiments across 0.6B-, 1.7B-, and 4B-parameter Qwen3 students show that ActFirst-OPD achieves average wall-clock training speedups of $2.3\times$ on ALFWorld, $1.8\times$ on WebShop, and $4.9\times$ on ScienceWorld over Vanilla OPD. It matches or exceeds all compared OPD baselines in mean task success rate across eight of nine benchmark-model settings. These results demonstrate that reasoning need not block acting during multi-turn agent distillation.

On-policy self-distillation (OPSD) trains a student to match a privileged teacher distribution along its own sampled trajectory. Standard OPSD applies this supervision to unverified student rollouts while conditioning the teacher on privileged context, typically a reference solution. We separate these roles in a factorial analysis and find that scaffold correctness has a stronger effect on downstream accuracy than context correctness. Unverified scaffolds create an imitation gap because the teacher can use information unavailable to the student. This gap shrinks with model scale, yet OPSD continues to supervise mostly unverified trajectories. In contrast, verified scaffolds remain effective even when the teacher is conditioned on the student's own unsuccessful rollout. Based on this finding, we introduce OASIS, which retains the OPSD objective but supervises mostly verified by label on-policy trajectories and replaces written solutions with unverified model-generated attempts as the teacher context. OASIS therefore requires only final-answer labels. Across Qwen3-1.7B, 4B, and 8B on AIME 2024, AIME 2025, and HMMT 2025, OASIS improves over the base model by 3.2--3.8 points on average, while OPSD's gain falls from 3.05 points at 1.7B to 0.14 at 8B. At 8B, OASIS improves over OPSD by 3.05 points, showing that verified on-policy scaffolds preserve the effectiveness of self-distillation as models scale.

On-policy distillation (OPD) trains a student to match the teacher's next-token distributions on the student's own trajectories and has yielded substantial empirical gains. Generalized variants allow the student to surpass the teacher by extrapolating an implicit reward in output space. The language-model head, however, attenuates this change anisotropically: much of the change encoded in the teacher's hidden states reaches the logits at a small fraction of its weight, and the sampled-token log-probability ratios on which output-space extrapolation relies inject noise that the extrapolation amplifies, making training unstable. We observe that reinforcement learning (RL) shifts a model's internal representations relative to its base checkpoint, and that the direction of this shift can be measured at every layer. Motivated by this observation, we propose RIDE (RL-Induced Direction Extrapolation), which extrapolates the RL-induced change directly in representation space: at every layer and token position, RIDE computes the residual between the teacher and its pre-RL checkpoint and regresses the student's hidden states toward targets displaced beyond the teacher along this residual. Conditioned on a sampled trajectory, this regression is equivalent to maximizing a linear directional reward defined by the residual under a quadratic penalty centered at the teacher, which makes explicit how the objective moves the student along the RL-induced direction while limiting its deviation from the teacher. Across four base/RL-teacher pairs spanning different scales, architectures, and pre-training lineages, RIDE approaches or exceeds the RL-trained teacher on every pair and is the only method whose mean does so, and it consistently outperforms output-space extrapolation, which degrades the student whenever the teacher is close to its base. Project page: https://github.com/xixixixixxxx/RIDE.

Language-model agents are usually trained by reinforcement learning from one reward per episode, and privileged self-distillation enriches it by letting the same policy, given a skill, teach its skill-free self through token probabilities. However, we identify two phenomena that question this channel. Invisible Advantage: a skill in context lifts WebShop success from 42.2% to 56.2%, yet changes the probabilities of fewer than a quarter of the sampled tokens. Much to Align: a skill changes the hidden states of over 80% of response tokens, in a way that linear probes can trace back to the specific skill. To exploit this, we propose Privileged Representation On-policy Self-Distillation (PR-OPD). After a GRPO warm start, the policy writes a hindsight skill for each trajectory, re-reads its own responses with that skill as a stop-gradient teacher, and aligns its projected hidden states to the teacher's at every layer alongside the reward objective, with no external skill library, separate teacher, or inference overhead. On ALFWorld and WebShop with two backbones, PR-OPD achieves the best overall results in every setting, improving over GRPO by up to 4.7 points in ALFWorld success and 14.0 points in WebShop accuracy. Code is available at https://github.com/balibata/PR-OPD.