2026-09-29
How to Loop MoE: Flatten the Experts, Untie the Attention
针对循环MoE模型如何循环的问题，提出Foil方法，通过展平专家和解耦注意力提升专家利用率。
Authors: Shouren Wang, Chuang Ma, Mohsen Hariri, Debargha Ganguly, Wang Yang, Xiaoqing Tong, Qianying Liu, Xiaotian Han, Vipin Chaudhary | Date: 2026-09-28
Link: http://arxiv.org/abs/2609.35751v1
Abstract
Looped Transformers reuse one block of layers several times: by spending extra computation they push a model of fixed size further, and so use its parameters more fully; while sparse mixture-of-experts (MoE) models activate only a few of many experts for each token. Looped MoE bridges these two design philosophies and gives MoE models new potential for better expert usage, but it raises a question: how to loop a MoE? We answer it with Foil. With the expert parameters and the expert compute per token held fixed, Foil (1) flattens the experts, halving the expert layers, doubling the experts per layer and doubling the passes, so that every routing decision chooses from a larger pool, and (2) unties the attention, giving each pass its own attention parameters while the experts and routers stay shared. Experiments show that Foil clearly outperforms the unflattened looped baseline: at 20B tokens every Foil model has lower pretraining loss than the baseline; at 100B tokens the loss improves monotonically with the degree of flattening, the most flattened Foil ending 0.012 nat below the baseline at equal parameters and compute, with downstream accuracy on par or better; untying the attention also yields more balanced and more confident routing at equal shape. Our ablations analyse why Foil works and turn the findings into design guidance for looped MoE: the returns of looping and of widening the expert layers amplify each other, routing confidence tracks healthy expert use better than load balance, and a sparse looped MoE should therefore use more experts per layer and more passes. Code and configurations are available at https://github.com/SR-A-W/how-to-loop-moe.KV-streams for Efficient Compaction in Agentic Reinforcement Learning
针对智能体强化学习中上下文压缩导致训练吞吐量低的问题，提出KV-streams策略，实现高效可训练的压缩。
Authors: Emiliano Penaloza, Dane Malenfant, Dheeraj Vattikonda, Roger Creus Castanyer, Siddarth Venkatraman, Abhay Puri, Jonathan Light, Matthew James Sargent, Augustine N. Mavor-Parker, Massimo Caccia, Lucas Caccia, Glen Berseth, Esmeralda S. Whitammer, Alessandro Sordoni, Minseon Kim, Marc-Alexandre Côté, Laurent Charlin, Guillaume Lajoie | Date: 2026-09-28
Link: http://arxiv.org/abs/2609.35750v1
Abstract
Scaling the horizon of agentic LLMs is bottlenecked by the need to fit ever longer context traces in GPU memory. Context compaction has been the most popular mechanism to alleviate this issue, keeping GPU memory constant for a given trace. Unfortunately, most compaction strategies rely on prefilling the LLM context many times over, hindering training throughput. To alleviate this bottleneck and enable efficient trainable compaction, we propose KV-streams, a plug-and-play strategy compatible with any compaction strategy that substantially increases throughput while showing no evidence of hindering performance. KV-streams enable scalable compaction by streaming the KV cache forward rather than flushing it after each compaction. We show that KV-streams enable three different compaction strategies, achieving a 2.6 to 5x wall-clock speedup in training. Beyond efficiency, we find that the streamed KV cache can act as a recurrent state, carrying forward information that has long since disappeared from the context. Specifically, in a controlled setting we show that, contrary to prior work, RL alone is all that is needed for this behavior to emerge. Overall, we show KV-streams to be an efficient and lightweight plug-and-play addition to any post-training pipeline.Improving Test-Time Scaling with Adaptive Looped Transformers
针对循环Transformer在测试时扩展性未被充分探索的问题，研究其准确率-计算斜率，发现循环能提升长输出生成的效率。
Authors: Yichen You, Tianyu Fu, Aosong Feng, Xingtai Lv, Xuefei Ning, Ning Ding, Yu Wang | Date: 2026-09-28
Link: http://arxiv.org/abs/2609.35748v1
Abstract
Looped transformers have demonstrated promising parameter efficiency by reusing layers for latent computation. Prior studies compare looped and non-looped models at matched parameters or per-token FLOPs. However, to the best of our knowledge, whether looping improves test-time scaling as outputs grow longer remains underexplored. Through post-training looped transformers, we study the accuracy-compute slope, measured as the accuracy gain per doubling of test-time decoding FLOPs. We find that existing looped transformers often yield steeper slopes than their non-looped baseline, yet underperform it at matched compute. While fixed-depth looping spends extra iterations on every token, our analysis shows that many tokens do not benefit from extra iterations. We therefore propose TaH2, which enables the model to focus extra iterations on the tokens that benefit from looping. It jointly post-trains the backbone and an iteration decider through lookahead depth supervision, which uses online labels indicating whether further iteration improves the prediction. TaH2 improves both the efficiency and attainable accuracy of test-time scaling. On challenging AIME benchmarks, TaH2 improves the accuracy-compute slope by 53% (2.74 vs. 1.79) over the non-looped baseline, exceeding the baseline's peak accuracy by about 3.4 points at matched test-time compute. As the maximum iteration depth increases, existing looped models largely plateau, while TaH2's gain over the non-looped baseline continues to grow from +2.8 points at depth 2 to +3.9 points at depth 8. Our code is available at https://github.com/thu-nics/TaH.The Hidden Perception Constraint in Task-Aware Compression
针对任务感知压缩中隐藏的感知约束问题，揭示并分析了压缩方案中自然存在的感知约束。
Authors: Sahan Liyanaarachchi, Semih Akkoc, Sennur Ulukus, Aylin Yener | Date: 2026-09-28
Link: http://arxiv.org/abs/2609.35684v1
Abstract
With the recent advancements of neural compressors, explicitly incorporating perception constraints into the design of compression schemes has gained significant attention. Traditionally, these perception constraints ensure that the distribution of the reconstruction does not significantly deviate from the distribution of the source, thus attesting to the perceptual quality of the reconstruction. In this work, we uncover several perception constraints that are naturally present in task-aware compression. In particular, we consider a problem where the primary task is reconstruction and the secondary task is classification (i.e., a statistical test). We study this problem at varying levels of domain information available to us and discuss how to utilize the naturally emerging perception constraints to design rate-minimal compression schemes that also maximize the utility of our secondary task. We show that in this setting, if the decision boundaries of the classifier are ill-defined (mismatch) for our source distribution, then matching onto a target distribution enhances our classification accuracy.MS-GLA: Multi-Scale Gated Linear Attention for Addressing Representational Bottlenecks via Multi-Temporal Resolution
针对门控线性注意力在单一时间分辨率下的表征瓶颈问题，提出多尺度门控线性注意力机制。
Authors: Prasoon Dev, Anirudh Sankar, Vasudeva Varma | Date: 2026-09-28
Link: http://arxiv.org/abs/2609.35664v1
Abstract
Gated Linear Attention (GLA) Transformers advance linear recurrent models through data-dependent gating, but face a core limitation: the fixed-capacity memory matrices across all heads operate at a single temporal resolution, where each token is processed individually, forcing them to simultaneously encode local syntactic patterns and long-range semantic structure, creating a representational bottleneck that gating alone is insufficient to resolve. We introduce Multi-Scale Gated Linear Attention (MS-GLA), which addresses this by distributing attention heads across multiple temporal resolutions. Coarser resolutions pool longer token spans naturally specializing toward long-range dependencies, while finer head groups retain sensitivity to local syntactic structure. A learnable, input-dependent fusion layer dynamically recombines head group outputs at each timestep, expanding effective memory capacity without increasing per-head state size. This multi-resolution decomposition draws on principles from Multi-Scale State-Space Models (MS-SSM), adapting them to the gated linear attention setting. We evaluate MS-GLA on language modeling, recall-intensive tasks, and long-context generalization. Across all settings, MS-GLA consistently achieves higher accuracy and lower perplexity than GLA at matched parameter counts, with up to 18.9% improvement on recall-intensive tasks and 9.5% lower average perplexity on language modeling benchmarks, validating multi-temporal resolution decomposition as a principled and effective extension of Gated Linear Attention.Cartridges++: KV Cache Compression without Off-Context Derailment
针对长文档KV缓存压缩导致上下文脱轨的问题，提出Cartridges++方法，避免压缩后的上下文信息丢失。
Authors: Sonia Laguna, Joao Monteiro, Marco Cuturi, Pierre Ablin, Eleonora Gualdoni | Date: 2026-09-28
Link: http://arxiv.org/abs/2609.35621v1
Abstract
Serving long documents to a Large Language Model (LLM) repeatedly is expensive: computations grow with context length, and the memory footprint of the key-value (KV) cache balloons. Compressed KV (CKV) representations aim to mimic the cache of a document and are typically computed once and for all, ahead of inference time. Methods to obtain CKVs range from drop mechanisms that reduce their number of columns, to learned approaches. Among the latter, Cartridges have emerged as a leading compression method, learning compact KV representations through distillation on relevant Q/A pairs. While existing evaluations focus primarily on whether Cartridges and other CKVs yield approximately similar responses to document-related, on-context queries, we investigate the crucial deployment question of whether they can handle off-context queries, something the native KV representation is particularly good at, thanks to the mechanics of attention. We observe a fundamental trade-off: while Cartridges perform better for on-context queries, heuristic-variants preserve better the original LLM's ability to operate off-context. We measure this through their capability to avoid context contamination in their response, retain general knowledge, and follow instructions. We propose Cartridges++, simple modifications to cartridges that retain off-context abilities at small or negligible cost. The router variant decides at inference time whether the query should use the learned long-context memory, while the data-mixing variant allocates a small fraction of training Q/As to queries outside the reference long document. Our study shows that assessing CKVs on document utility alone can mask substantial degradation in broader model capabilities, yet those issues can be fixed with benign changes to CKV inference or training.Output-aware Residual Stream Pruning for Large Language Models
针对残差流剪枝忽略下游层敏感性的问题，提出输出感知的敏感性剪枝方法，提升剪枝后模型性能。
Authors: Chayne Thrash, Kevin Chen, Soheil Kolouri | Date: 2026-09-28
Link: http://arxiv.org/abs/2609.35579v1
Abstract
Residual stream pruning methods reduce inference cost by shrinking the model's hidden dimension, but existing approaches typically choose these dimensions by minimizing activation reconstruction error. This criterion implicitly treats all perturbation directions as equally important, ignoring the sensitivity of downstream layers. We introduce a sensitivity-aware approach to residual-stream pruning that directly accounts for this direction-dependent sensitivity. Using a second-order approximation to the output KL divergence, we characterize the effect of a residual-stream perturbation through both its activation covariance and the local sensitivity of the model output. The resulting subspace selection objective couples these two quantities, but is difficult to optimize directly. We derive a tractable spectral upper bound that reduces subspace selection to an eigendecomposition of a sensitivity-weighted covariance matrix, retaining the efficiency and structural simplicity of rotation-based pruning methods. Across several instruction-tuned language model families, our method consistently reduces calibration KL divergence relative to activation-only pruning and improves perplexity and downstream task performance over a range of compression levels. Our results show that preserving activation energy alone is insufficient for residual-stream pruning, and that explicitly accounting for how perturbations propagate to the model output provides a more effective criterion for selecting dimensions to remove.FactorEngram: Factorized N-gram Memory with Basis-Level Gating for Language Models
针对基于查找的记忆中多义模式无法选择性读取的问题，提出分解式N-gram记忆与基础级门控机制。
Authors: Bowen Yang, Jingbo Zhou, Qinghong Miao, Hua Wu | Date: 2026-09-28
Link: http://arxiv.org/abs/2609.35578v1
Abstract
Lookup-based memory has been a promising way to scale the parameters of large language models (LLMs). It retrieves learned representations of local token patterns, such as n-grams, instead of reconstructing them through successive layers of computation. However, existing designs such as Engram treat each retrieved embedding as a monolithic unit. Each embedding is stored in its own hashed slot and modulated by a single scalar gate. As a result, polysemous patterns cannot selectively read out the components of their memory that are relevant to the context. Moreover, parameters are shared only through hash collisions, which are largely unrelated to semantics. We propose FactorEngram, a factorized n-gram memory with basis-level contextual gating. FactorEngram retrieves sparsity-regularized coefficients over a dictionary of basis vectors shared across patterns, so related patterns can reuse common components. The same dictionary is also used for gating. The backbone hidden state is scored against each basis vector to gate the corresponding coefficient before reconstruction, which lets the context modulate each memory component individually. FactorEngram also covers both individual tokens and multi-token n-grams, and we systematically study where the memory branch should be inserted. On 340M- and 1B-parameter Transformer backbones, FactorEngram improves language modeling and downstream task performance. Ablation studies confirm the contribution of each component and identify insertion before the attention sublayer in the middle layers as an effective configuration.Continuous Context Management
针对长时智能体累积完整交互历史导致内存开销大的问题，提出连续上下文管理，在每轮进行压缩。
Authors: William Hoy, Jingxuan Fan, Nurcin Celik, Xu Pan | Date: 2026-09-28
Link: http://arxiv.org/abs/2609.35540v1
Abstract
Long-horizon large language model (LLM) agents commonly retain their complete interaction history until compaction is triggered at a predefined threshold. We study Continuous Context Management (CCM), which performs compaction at every turn to prevent interaction history from accumulating in the active prompt. At each turn, a CCM agent emits an updated memory together with an environment action; its next prompt contains the original task, retained memory, and newest observation rather than the complete transcript. We first evaluate CCM without fine-tuning on TerminalBench-2 using Claude Sonnet 4.6, Claude Opus 4.6, GLM-5, and Kimi K3. CCM substantially reduces cumulative input usage and active-prompt size, although it lowers task success for most models while preserving performance for Kimi K3. We use GRPO with privileged full-history distillation to improve CCM in open-weight models. A frozen copy of the student's initial model scores each sampled student action under the complete history reconstructed from that student's rollout, providing dense action-token supervision without a separate teacher rollout or reference solution. On WebShop, this objective substantially improves CCM over GRPO at both evaluated model scales and surpasses full-history GRPO for Qwen3-4B-Instruct, though not for Qwen3-8B. On Endless Terminals, the augmented method provides a modest improvement over GRPO, with both CCM policies outperforming the untrained full-history baseline. These results demonstrate that CCM is a viable inference paradigm for agents operating with substantially reduced retained context and that its performance can be improved through reinforcement learning with privileged full-history distillation.Beyond Token Scale: Chunk-Level Sparse Autoencoders for Reliable Semantic Feature Discovery
针对稀疏自编码器在令牌级目标下难以发现可靠语义特征的问题，提出块级稀疏自编码器。
Authors: Xu Wang, Yifan Yang, TingHao YU, Difan Zou | Date: 2026-09-28
Link: http://arxiv.org/abs/2609.35521v1
Abstract
Sparse autoencoders (SAEs) expose features that help us understand and steer language models, but faithful reconstruction does not guarantee informative concepts. Token-level objectives reward lexical and formatting details alongside semantic content, all competing for a limited sparse budget. We introduce a family of chunk-level SAEs that encode mean-pooled activations over chunks, each a contiguous span of tokens: Mean-Chunk reconstructs the observed chunk, Cross-Chunk predicts an independently processed neighbor, and Joint-Chunk combines both targets. These designs separate the effect of a larger observation unit from that of predicting information shared across passages. With matched training data, chunk-level SAEs remain powerful interpretability tools while learning reliable semantic features that capture high-level concepts and respond selectively to relevant content. Their strengths are complementary: Mean-Chunk improves high-level feature discovery, reasoning detection beyond surface cues, and steering; Cross-Chunk leads document retrieval and classification transfer while producing selective, persistent features. Changing what an SAE sees and predicts yields reliable semantic features for more meaningful tasks. We demonstrate their practical value through gains across downstream tasks such as retrieval, reasoning detection, and steering.Reward-Aligned Reweighting for On-Policy Distillation
针对在策略蒸馏中均匀加权忽略后续推理价值的问题，提出奖励对齐的重加权方法。
Authors: Haofeng Xu, Junwei Su, Lansong Diao, Wenchao Zhou, Chuan Wu | Date: 2026-09-28
Link: http://arxiv.org/abs/2609.35517v1
Abstract
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.An RL View of OPD: Least Square Policy Distillation for Sample-Efficient LLM Reasoning
针对在策略蒸馏样本效率低的问题，通过强化学习视角提出最小二乘策略蒸馏框架，引入探索和数据重用。
Authors: Shangzhe Li, Yuxiao Yang, Tianrun Yu, Kaixiang Zhao, Xiaoyun Wang, Taylor W. Killian, Weitong Zhang | Date: 2026-09-28
Link: http://arxiv.org/abs/2609.35505v1
Abstract
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.SOLO: Pretraining Billion-Parameter Language Models with Shared-Output Local Learning
针对本地学习无法扩展到十亿参数预训练的问题，提出共享输出的本地学习方法SOLO。
Authors: Bojian Yin, Shurong Wang, Yuqi Pan, Guoqi Li | Date: 2026-09-28
Link: http://arxiv.org/abs/2609.35440v1
Abstract
Large language models are trained with backpropagation, whose global gradient coordinates all layers but forces each to hold its activations and wait for the gradient to pass back through every deeper layer. Conventional local learning removes this update locking by training each module to predict the target through its own readout, but has not scaled to billion-parameter pretraining. We identify these private readouts as a key weakness, since they leave each module without information from deeper modules. We propose Shared-Output LOcal learning (SOLO), which replaces them with a shared, read-only copy of the final module's readout, the only one trained on the output of the whole network. Taken from the previous step, the copy transmits information from the final module without passing gradients between modules or reintroducing update locking. SOLO approaches backpropagation on Transformers of 340M to 2B parameters pretrained on 15B tokens, staying within one point in average zero-shot accuracy with a perplexity gap that narrows with scale. Readout ablations attribute SOLO's improvement over private readouts to sharing. Without update locking, each of p pipeline stages holds activations for O(1) micro-batches instead of O(p). The freed memory permits larger micro-batches, which reach up to 1.44x the best measured throughput of pipeline backpropagation on the same partition. To our knowledge, SOLO is the first local learning method to show such memory and throughput gains in billion-parameter language-model pretraining. Local learning thus becomes a practical alternative to backpropagation for large-scale pretraining.Frontier Learning: Training LLM Reasoners at the Edge of Capability
针对基于强化学习的后训练在固定问题池上学习信号过时的问题，提出前沿学习方法，动态生成训练问题。
Authors: Robin Faro, Shyam Sundhar Ramesh, Ilija Bogunovic, Aurelien Lucchi | Date: 2026-09-28
Link: http://arxiv.org/abs/2609.35426v1
Abstract
Reinforcement Learning-based post-training of Large Language Models (LLM) has been successfully applied to improve their reasoning capabilities. Existing pipelines primarily finetune LLMs on a fixed pool of problems specified prior to training using the GRPO loss. This is fundamentally limiting, as learning signal arises only when policy rollouts mix successes and failures, causing the useful portion of any fixed pool to quickly become stale as the model improves. To address this, we propose frontier learning, an open-ended post-training approach in which procedural generators are used online to continually produce informative training problems. It treats the generator's task-specific parameters as a search space and uses a regret signal to prioritize and explore frontier difficulty levels in order to focus training at the edge of the model's evolving reasoning capabilities. Across several reasoning tasks and model families, our approach consistently achieves higher relative gains over fixed-pool baselines, demonstrating that effective post-training requires not only selecting useful problems, but continually generating them at the edge of capability.Fiona: Accelerating FHE Inference with Packing-Aware Ternary Weights
针对全同态加密推理速度慢的问题，提出感知打包的三元量化方法Fiona，加速加密推理。
Authors: Yiteng Peng, Zhibo Liu, Dongwei Xiao, Shuai Wang | Date: 2026-09-28
Link: http://arxiv.org/abs/2609.35352v1
Abstract
Fully homomorphic encryption (FHE) enables neural network inference directly on encrypted inputs, but it remains orders of magnitude slower than plaintext in- ference. Applying the server's plaintext weights to encrypted activations involves plaintext-ciphertext multiplications (PMult) and accounts for more than half of inference time in recent systems. Ternary quantization can replace these multipli- cations with additions and subtractions, but the savings rarely materialize under packed execution. A single PMult applies a weight group fixed by the packing layout and can be avoided only when all its weights share the same ternary value. Ternarizing all groups, however, largely degrades accuracy. We present FIONA, an offline optimizer that selectively ternarizes weights within a given packing layout based on the estimated effect of ternary conversion on the model's performance. FIONA encourages a shared ternary value within each weight group and retains full-precision weights for sensitive groups, so ternar- ized and full-precision paths coexist within a layer. It then compiles these hybrid operators exactly, applying common scaling factors once to accumulated inputs and reusing sums across outputs. Weight ternarization can also narrow the input ranges of downstream polynomials. FIONA fits lower-degree replacements under a cumulative accuracy budget, reducing multiplicative depth and bootstrapping. On VGG11, ViT, and BERT, FIONA reduces PMult operations by 53.4-79.5% and accelerates end-to-end encrypted inference by 2.38x, 1.68x, and 1.84x, re- spectively, with less than 1% accuracy loss across all three models.Quasi Linear Kernel Attention with Infinite Capacity
针对核注意力在保持表达能力与线性计算间权衡的问题，提出具有无限容量的准线性核注意力。
Authors: Nicolaj Rux, Johannes Hertrich, Sebastian Neumayer | Date: 2026-09-28
Link: http://arxiv.org/abs/2609.35349v1
Abstract
The evaluation cost of transformers with softmax attention scales quadratically with sequence length. Kernel attention addresses this by replacing softmax with a more general kernel function. In this paper, we aim to identify kernels that retain the expressivity of attention while enabling quasi linear computation. To quantify expressivity, we introduce a capacity for each kernel, measuring the maximum sequence length for which the attention matrix can approximate the identity. A higher capacity thus indicates greater expressivity. We show that expressive kernels like softmax, Gauss, and Laplace have infinite capacity. In contrast, common quasi linear kernels, such as those derived from finite dimensional feature maps, exhibit finite capacity. As a solution, we propose additive kernels constructed from univariate spline and polynomial exponential kernels. We prove that these maintain infinite capacity while allowing quasi linear computation via sorting. Finally, we implement additive sorting kernels efficiently and benchmark them against modern softmax backends, demonstrating advantages for long sequences.Beyond Teacher Assignment: Domain-Normalized Multi-Teacher On-Policy Distillation
针对多教师在策略蒸馏中路由决策不考虑特征强度的问题，提出领域归一化的多教师蒸馏方法。
Authors: Xin Li, Hao Jiang, Xin Gao, Annan Wang, Yuchen Xie, Jinghao Guo, Xingwei Qu, Yichi Zhang, Chau Yuen | Date: 2026-09-28
Link: http://arxiv.org/abs/2609.35347v1
Abstract
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.LionMuon: Alternating Spectral and Sign Descent for Efficient Training
针对语言模型预训练优化器效率问题，提出LionMuon，交替使用谱下降和符号下降以提升训练效率。
Authors: Arman Bolatov, Artem Riabinin, Nikita Kornilov, Andrey Veprikov, Samuel Horváth, Martin Takáč, Aleksandr Beznosikov | Date: 2026-09-28
Link: http://arxiv.org/abs/2609.35297v1
Abstract
Pretraining a language model takes enormous compute, and the right optimizer can save a good part of it. Muon's spectral step gives a stronger direction than a sign step, but it is expensive. Every step runs Newton-Schulz iterations on the full matrix and, in distributed training, an extra all-reduce. Sign steps, as in Lion and Signum, are cheap and stay local to each device. We propose LionMuon, which takes one Muon step every $P$ iterations and Lion steps in between, with a single dual-EMA momentum buffer shared by both. Muon's compute and communication are paid once per $P$ steps, and the optimizer state is half of AdamW's. A single-EMA variant, SignMuon, already improves on Muon. We prove complexity bounds under heavy-tailed noise in which the period sets an interpolation between Muon's and Lion's smoothness and noise constants, and which say when LionMuon is faster than both. On 124M and 355M models trained on FineWeb, LionMuon with $P=2$ and $P=5$ reaches a lower loss than Muon, AdamW, Lion and Signum at the same number of tokens. Under 4-GPU data-parallel training it reaches Muon's final loss with a third less wall-clock on PCIe, and it beats the communication-efficient Muon variants Dion and MuonBP on loss at no more exposed communication, while keeping the exact gradient. Code: https://github.com/brain-lab-research/lion-muonWavePP: High-Throughput Pipeline Parallel LLM Prefill under Prefix Reuse
针对流水线并行在前缀重用下调度开销大、吞吐量低的问题，提出WavePP高吞吐量流水线并行预填充。
Authors: Aaryam Sharma | Date: 2026-09-28
Link: http://arxiv.org/abs/2609.35263v1
Abstract
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.Rubric-Aware On-Policy Self-Distillation for LLM Personalization
针对LLM个性化中评分标准利用粒度粗的问题，提出评分标准感知的在策略自蒸馏方法。
Authors: Yilun Qiu, Xiaoyan Zhao, Chengbing Wang, Cilin Yan, Rui Zu, Wanyang Zhang, Xiaolong Jiang, Jiayin Cai, Yang Zhang | Date: 2026-09-28
Link: http://arxiv.org/abs/2609.35262v1
Abstract
LLM personalization aims to generate responses aligned with individual users' preferences and needs. User-specific rubrics make these expectations explicit, providing direct supervision on what a satisfactory answer should cover. Existing rubric-guided approaches, however, exploit such guidance only at a coarse granularity, either by using rubrics to supervise the prediction of relevant aspects for subsequent generation or by reducing aspect coverage to a single response-level reward for reinforcement learning. This leaves a gap between specifying what a personalized answer should contain and teaching the model how to generate it. To bridge this gap, we propose GRASP, a rubric-aware on-policy self-distillation framework for LLM personalization that turns user-specific rubric aspects into fine-grained, token-level supervision. Specifically, GRASP pairs a rubric-free student with a rubric-informed teacher that additionally receives the target user-specific rubrics. By aligning their next-token distributions along on-policy trajectories generated by the student, GRASP transfers the teacher's rubric-conditioned guidance into the student, translating user-specific semantic requirements into dense token-level supervision. Since rubric-informed teachers can still produce inadequate supervision, we further introduce Rubric-based Teacher Validation (RTV), which retains only instances where the teacher sufficiently covers the target aspects, improving both supervision quality and training efficiency. Experiments on the LaMP-QA benchmark for personalized question answering demonstrate that GRASP achieves state-of-the-art performance across multiple backbones, supporting the effectiveness of rubric-guided token-level supervision for personalization. To ensure reproducibility, our code is available at https://github.com/SnowCharmQ/GRASP.Understanding On-Policy Distillation: A Mechanistic Interpretability Perspective via Sparse Crosscoders
针对在策略蒸馏实际蒸馏内容不明确的问题，通过稀疏交叉编码器从机制可解释性角度进行分析。
Authors: Zichao Yu, Qianshuo Ye, Xu Wang, Difan Zou | Date: 2026-09-28
Link: http://arxiv.org/abs/2609.35210v1
Abstract
On-policy distillation (OPD) is a widely adopted post-training technique for LLM reasoning. It is commonly believed to transfer knowledge from a stronger teacher, yet what OPD actually distills into the student's internal representations remains unclear. We study this question with sparse crosscoders, which learn one feature dictionary shared by the student before and after OPD and the teacher. Standard crosscoder analyses, however, identify model-specific features but cannot tell how a model's use of its features changes, since all models are encoded into one set of feature activations. We therefore propose the swap readout, which reads each student checkpoint's feature activations on its own, measuring how training changes the student's use of each feature, even for checkpoints unseen by the crosscoder. Across three OPD settings, we find that OPD neither creates features nor passes on the teacher's own, and leaves the firing rates of over 98% of the student's frequently used features within 20%. We further examine the SFT warm-up on the teacher's rollouts that commonly precedes OPD and makes it more effective. Rather than adding features, the warm-up reweights the shared ones in two ways. First, it already raises and lowers many of the features that OPD later raises and lowers, doing part of OPD's work in advance. Second, it changes features that OPD alone would not, notably those for conversation format, reasoning style, and mathematical notation, and these changes persist through OPD. Imposing this reweighting on a directly distilled student's features, without changing its weights, brings its accuracy close to that of the warmed-up student, whereas the same change on shuffled features does not. Together, these findings suggest that OPD reweights existing features rather than acquiring new ones: the student learns from the teacher how to use the features they already share.Beneath the Tokens: A Performance Engineering Study of Multi-Token Prediction in GPU-Accelerated LLM Inference
针对自回归LLM推理受内存移动和顺序执行限制的问题，评估多令牌预测在GPU加速推理中的性能。
Authors: Suwesh Prasad Sah | Date: 2026-09-28
Link: http://arxiv.org/abs/2609.35188v1
Abstract
Autoregressive large language model inference repeatedly invokes the target model to generate one token at a time, making generation sensitive to GPU memory movement and sequential execution. This study evaluates two-token multi-token prediction (MTP) against autoregressive decoding in a controlled single-request deployment on an NVIDIA A10G GPU. A 360-request benchmark covered plain-text, reasoning-intensive, and tool-calling workloads, while runtime telemetry, Nsight Systems, PyTorch Profiler, and selected Nsight Compute measurements were used to explain the observed performance. MTP increased output throughput by \(1.91\times\) to \(2.19\times\) across all prompts and reduced time to first output by 10.0--14.2\%. Median mean acceptance length ranged from 2.370 to 2.595 tokens per verification iteration. Profiling showed that MTP introduced a longer and more complex execution path, including proposal, sampling, attention, gathering, and reduction operations. However, it required 56.4--78.1\% fewer executions of the selected repeating CUDA Graph per generated token. The dominant MTP GEMM kernel was not faster than the dominant autoregressive GEMV kernel, and selected instances of both approached the A10G memory-bandwidth limit. These results show that MTP improved inference through amortization: greater token progress reduced repeated GPU execution sufficiently to outweigh the additional speculative-execution cost.EdgeCraft: Automated Model Crafting for Edge IoT
针对边缘物联网ML模型制作流程碎片化、难以扩展的问题，提出LLM驱动的自动化模型制作系统EdgeCraft。
Authors: Genglin Wang, Kaiwei Liu, Liekang Zeng, Wangsong Yin, Shangcheng Jin, Guoliang Xing, Zhenyu Yan | Date: 2026-09-28
Link: http://arxiv.org/abs/2609.35167v1
Abstract
Machine learning (ML) increasingly powers Internet of Things (IoT) applications at the edge. Yet producing a deployable edge ML artifact for a specific scenario requires navigating a huge search space spanning data representation, model design, training on domain-specific data, and runtime customization. This workflow is fragmented and difficult to scale across diverse edge applications. We present EdgeCraft, an LLM-driven system that turns high-level intent into deployable edge ML artifacts. Building such a system raises two challenges: (1) How can an LLM be guided to find high-quality solutions that meet dynamic SLOs for task quality, latency, and energy? (2) How can trustworthy target-device verification be obtained at low cost? EdgeCraft addresses these challenges with two designs. (1) A constraint-aware synthesis tree explores alternative candidates and uses measured SLO gaps to guide each improvement. (2) A multi-fidelity verifier progressively combines low-cost checks with full target-device verification to reduce verification cost while preserving reliable verification results. It also records verified failures for reuse, avoiding repeated device work. To support concurrency, EdgeCraft provides a multi-tenant runtime that runs cloud training and target-device verification in parallel while isolating requests. Across 50 public tasks, EdgeCraft exceeds the task-specific Reference in best-observed quality on 40 tasks and finds an SLO-feasible artifact on 45, with the two outcomes overlapping on 38 tasks. Moreover, EdgeCraft achieves competitive performance on our self-collected SEN dataset, suggesting its generalizability to real-world IoT sensing tasks.PEARL: Adaptive Prefill-Decode Execution with Elasticity for Agentic Reinforcement Learning
针对智能体强化学习中多轮 rollout 成本高、资源利用不均的问题，提出自适应预填充-解码执行框架PEARL。
Authors: Jiaan Zhu, Wei Gao, Youhui Bai, Zewen Jin, Ju Huang, Siran Yang, Jiamang Wang, Lin Qu, Cheng Li | Date: 2026-09-28
Link: http://arxiv.org/abs/2609.35158v1
Abstract
Multi-turn rollout dominates the cost of agentic reinforcement learning (RL). Asynchronous execution and elastic GPU resources can accelerate this stage, but adding rollout replicas yields diminishing returns while training GPUs remain idle between updates. We observe that effective resource use also depends on the prefill--decode (PD) configuration. Both the choice between colocation and disaggregation and the optimal PD ratio vary with the workload, making resource scaling and PD configuration interdependent. Exploiting this opportunity requires selecting effective configurations and realizing their benefits within transient resource-availability windows despite reconfiguration costs. We present PEARL, an asynchronous agentic RL system that coordinates external resource elasticity, temporary reuse of idle training GPUs, and adaptive PD execution. PEARL maintains a unified GPU--worker--role state and uses runtime profiles to predict rollout batch completion time, accounting for environment-induced reductions in decode concurrency. It selects the PD mode and ratio under the current GPU budget and translates each decision into an incremental transition plan that minimizes worker and role changes. Cost-aware switching and borrowing policies suppress transitions with insufficient expected benefit while ensuring timely return of training GPUs. Our evaluation show that PEARL achieves $2.17$--$2.79\times$ the throughput of fixed-resource ROLL across different LLMs. Compared with RLBoost+, throughput improves by up to approximately 26.9\% for Qwen3-8B and 36.3\% for Qwen3-30B-A3B.AReaL-TIK: Stateful Agentic Optimization of Unified RL Kernels through an Optimization IR
针对强化学习后训练中不同GPU内核导致数值不一致的问题，提出通过优化IR进行有状态的统一内核优化。
Authors: Ran Yan, Youhe Jiang, Jiayi Nie, Wenshuang Li, Yingqi Peng, Taiyi Wang, Tongkai Yang, Binhang Yuan | Date: 2026-09-28
Link: http://arxiv.org/abs/2609.35140v1
Abstract
Reinforcement learning (RL) post-training often uses distinct GPU kernels for rollout and policy update. In synchronous PPO and GRPO, numerical disagreement can perturb ratios between current token probabilities and those assigned during rollout. Recomputing rollout log-probabilities with the policy-update backend avoids this discrepancy but adds a forward pass. Bitwise-consistent unified kernels permit reuse when the policy snapshot and probability processing match the objective. Their optimization must preserve agreement across distinct execution regimes. We present KernelBraid, an agentic framework starting from a hand-tuned, bitwise-consistent implementation. Its optimization intermediate representation (IR) organizes source-code search by linking implementations and modifications to numerical requirements, workload measurements, and derivation history. The agent coordinates changes and retains verified intermediates for further exploration; promotion requires passing correctness checks and improving aggregate latency within per-workload limits. Across 12 end-to-end training configurations on H20, KernelBraid achieves 1.10x average throughput relative to AReaL with log-probability recomputation, and the mean training-reward ratio rounds to 1.00x. Isolated-layer profiling yields 1.40x average speedup in summed phase time across 15 model-GPU pairs. Operator-level evaluation covers correctness and performance for 10 operators on A100, H20, and H200, all passing the prescribed bitwise checks. Unified-attention search achieves 2.52x speedup in summed workload latency over the starting implementation using 7M LLM tokens; ablations assess the contributions of retained evidence and branch exploration to search efficiency and attained performance. Our code is open-sourced at https://github.com/areal-project/AReaL-TIK.CacheRepair: Learning to Repair Cross-Chunk Context in RAG for KV Cache Fusion
针对多文档RAG中预计算KV缓存缺乏跨块注意力信息的问题，提出CacheRepair学习修复跨块上下文。
Authors: Genglin Wang, Wangsong Yin, Yeerzhati Abudunuer, Haoxuan Xu, Guoliang Xing, Zhenyu Yan | Date: 2026-09-28
Link: http://arxiv.org/abs/2609.35139v1
Abstract
Multi-document retrieval-augmented generation (RAG) requires a language model to process multiple retrieved text chunks before answering a question. Precomputing each chunk's KV cache independently and concatenating the caches when the chunks are retrieved can accelerate this step. However, the assembled cache lacks cross-chunk attention information, reducing answer quality. Selective recomputation methods recover the missing cross-chunk context by rerunning the target LLM on selected tokens, incurring substantial online computation. We introduce CacheRepair, a lightweight network that learns the difference between independently computed KV caches and those produced by processing the chunks together. The network combines compressed KV features with token embeddings and uses attention that is bidirectional within each chunk and flows from earlier to later chunks. Each repair block receives the compressed cache features, and the predicted residual is added to every document token's cache. Each repair network is trained for a specific frozen target LLM on a generic retrieval corpus and reused across downstream datasets. Our analysis shows that repair reduces KV errors both near chunk boundaries and throughout chunk interiors. Evaluation across three target LLMs and four downstream datasets places CacheRepair on the measured answer-quality-latency Pareto frontier in eleven of twelve model-dataset combinations. Reported time to first token (TTFT) includes online cache transfer and repair. Across all twelve combinations, the largest repairers achieve 1.69-4.61$\times$ speedups in median TTFT over full prefill and improve mean F1 by 2.1-26.1 percentage points over direct cache reuse.TempoKV: Timely Staging of LLM KV Caches for Memory-Semantic Flash
针对LLM服务中可重用KV缓存超出GPU内存的问题，提出TempoKV，基于时间感知的分层缓存管理。
Authors: Jay H. Park, Hyungjun Kim, Dong Kim | Date: 2026-09-28
Link: http://arxiv.org/abs/2609.35065v1
Abstract
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%.SPIDER: Multi-Layer Semantic Token Pruning and Adaptive Sub-Layer Skipping in Multimodal Large Language Models
针对多模态大语言模型中数据与计算冗余耦合的问题，提出多层语义令牌剪枝和自适应子层跳过方法SPIDER。
Authors: Tianxiang Chen, Zhentao Tan, Zi Ye, Yue Wu, Xiaobing Tu, Jinkui Ren, Xiantao Zhang, Tao Gong, Qi Chu, Nenghai Yu, Xipeng Qiu, Jieping Ye | Date: 2026-09-28
Link: http://arxiv.org/abs/2609.34977v1
Abstract
Multimodal Large Language Models face significant efficiency challenges that stem from two distinct yet coupled sources: data redundancy and computational redundancy. While most methods focus on data redundancy by pruning visual tokens from the output of the visual encoder or computing redundancy in LLM decoders using blockwise importance, the finer-grained inter-layer representation shifts and the distribution differences within the layers themselves have not been fully explored. In this work, we comprehensively investigate this dual-level inefficiency. We posit that intermediate layer tokens from vision encoders should be considered for effective visual token pruning, as semantic focus shifts across layers, with middle-layer tokens capturing more detailed object-centric information that deeper layers may abstract away. Furthermore, we reveal the differential contributions of Attention and FFNs across distinct LLM decoder layers. Building upon these discoveries, we propose \textbf{SPIDER}, a training-free framework that integrates multi-layer \underline{\textbf{S}}emantic visual token \underline{\textbf{P}}run\underline{\textbf{I}}ng with an a\underline{\textbf{D}}aptive sub-lay\underline{\textbf{ER}} skipping mechanism. Experimental evaluations demonstrate that SPIDER consistently maintains strong performance across various MLLM architectures and reduction ratios. For instance, on LLaVA-NeXT-7B, SPIDER reduces FLOPs by $79\%$ while maintaining 96$\%$ of the baseline performance.Just MLPs: Efficient Visual State Reconstruction for Multimodal Language Models
针对多模态大语言模型中长视觉令牌序列计算开销大的问题，提出仅用MLP进行高效视觉状态重建。
Authors: Jingdi lei, Junxian Li, Di Zhang, Zhanqiu Zhang, Yiwen Guo, Soujanya Poria | Date: 2026-09-28
Link: http://arxiv.org/abs/2609.34972v1