2026-09-11
Building py-kvcache: A Performance Characterization of External KV Caching for vLLM with NVMe SSDs
针对vLLM外部KV缓存性能权衡问题，通过多层级特性分析，揭示了缓存性能依赖于传输粒度和内存使用等因素。
Authors: Joseph Kanichai, Tiziano De Matteis, Animesh Trivedi | Date: 2026-09-10
Link: http://arxiv.org/abs/2609.11744v1
Abstract
Prefix caching can reduce the time to first token (TTFT) of long-context LLM requests by reusing previously computed key-value (KV) states, but for short prefixes or fast GPUs, recomputation can be faster than loading from an external cache. We characterize this tradeoff in vLLM across GPU, CPU, and NVMe tiers using synthetic workloads, long-context benchmarks, production traces, and find that cache performance depends on transfer granularity, intermediate memory use, and when transfers enter the request schedule, not only on device bandwidth. These findings motivate py-kvcache, a vLLM KV Offload connector with asynchronous direct I/O, bounded shared staging, and scheduler-aware preloading, which starts disk reads while requests are still waiting, overlapping with compute. At 80k tokens, py-kvcache loading from disk is 2.0x faster than LMCache, with preloading contributing 1.34x. With GPU, CPU, and disk caching enabled, it is 1.23x faster than LMCache and within approximately 4% of the native vLLM KV Offload implementation. LongBench and SCBench show that these benefits extend to irregular prefix chains and multi-turn workloads. Bailian trace replays improve TTFT on a weaker GPU, but on an H100 the average request falls below the break-even point and GPU memory alone retains enough prefixes. External KV caching should therefore be treated as a setup specific admission decision. The py-kvcacheimplementation is available at: https://github.com/atlarge-research/py-kvcache.LOCUS: Task-Aware Low-Rank Post-Training for Token-Efficient Language Generation
针对LLM服务成本随输出长度增长问题，提出任务感知低秩后训练方法LOCUS，在保持效用下最小化输出token成本。
Authors: Dongfang Zhao | Date: 2026-09-10
Link: http://arxiv.org/abs/2609.11739v1
Abstract
Large language model serving costs scale directly with output sequence length, yet standard preference alignment often inflates response verbosity without improving utility. We study whether the parameterization of post-training updates affects generation length: low-rank subspaces alter sequence length without modifying the alignment loss. We present LOCUS, a method that selects a task-aware low-rank adaptation subspace to minimize output-token cost subject to a utility constraint. Within this subspace, post-training retains the native preference objective with a frozen backbone. Across Anthropic HH-RLHF dialogue preferences, we evaluate two $\sim$3B decoder backbones, Pythia-2.8B and Qwen2.5-3B, against protocol-matched full-parameter DPO and DrDPO branches and the released SamPO checkpoint. LOCUS reduces continuation length by up to 39.84\% on Pythia-2.8B and by 14.87--17.58\% on Qwen2.5-3B while updating only 0.24--0.28\% of model parameters, with no material change in the internal preference diagnostic.Why Does Post-Training Quantization Work?
探究训练后量化为何有效，发现量化模型能抑制隐藏状态误差累积，从而保持下游任务性能。
Authors: Yuxiang Chen, Michael Beyer, Jun Zhu, Jianfei Chen | Date: 2026-09-10
Link: http://arxiv.org/abs/2609.11716v1
Abstract
Post-training quantization compresses large language models (LLMs) by storing their weights at reduced precision, and each quantized weight introduces an error into the hidden states. Naively, these errors should accumulate with depth and corrupt next-token prediction; randomly initialized models accumulate these discrepancies rapidly, whereas quantized pretrained models accumulate much less hidden-state error and largely maintain downstream task performance, even though they were never trained with quantization noise. This raises the question we address: why does post-training quantization work? Comparing full-precision and quantized forward passes, we identify two mechanisms that characterize pretrained quantization robustness. First, the error a layer newly introduces tends to oppose the error it inherits from the layer's input. The two cancel partially such that the discrepancy between full-precision and quantized passes grows slowly. This counteracting residual interaction develops during pretraining. Our quantitative analysis identifies it as a major factor slowing hidden-error growth. Second, LM-head geometry preferentially preserves the scores and probabilities of high-ranked tokens, which typically represent the model's most confident predictions. Together, these mechanisms explain why quantization error that passes through numerous layers can still produce only small output changes, and we verify the findings across models and quantization settings.Structured Transforms for Low-Overhead Quantization of Language Models
针对LLM量化开销问题，提出基于符号随机化DCT的结构化正交变换，提升Kashin分解量化算法的收敛性和效率。
Authors: Daria Cherniuk, Alexander Rudikov, Boris Kashin, Ivan Oseledets | Date: 2026-09-10
Link: http://arxiv.org/abs/2609.11687v1
Abstract
We revisit Kashin-decomposition-based weight quantization for large language models and propose an improved algorithm with stronger convergence properties and structured, efficient orthogonal transforms. The method retains the core factorization of each weight into two components -- one with bounded infinity norm and the other with bounded infinity norm after an orthogonal transformation -- but replaces the dense random orthogonal matrix with a sign-randomized Discrete Cosine Transform (DCT), reducing the per-iteration cost from $\mathcal{O}(N^2)$ to $\mathcal{O}(N \log N)$. The proposed greedy algorithm with alternating updates guarantees the four-peak distribution required for stable 2-bit clustering of each factor and admits closed-form initialization of cluster centers, removing the multi-restart k-means bottleneck of prior work. Composed with OPTQ-style sequential error compensation and QuIP-style incoherence preprocessing, the resulting JAX pipeline is competitive with OPTQ, QuIP, QuIP-RG and a fine-tuning- and vector-quantization-free variant of QuIP# at 4-bit per channel on OPT, Llama-2 and Pythia, with favorable wall-clock scaling. The bounded-$\ell_\infty$ factorization is also notably robust: on stress configurations where QuIP variants diverge to four-digit perplexity (Pythia-6.9B) or abort with NaNs in LDL back-substitution (Mistral-7B), Kashin-DCT remains numerically stable and stays close to FP16 baseline. At inference time, each weight decomposes into two 2-bit factor codes per channel that are structurally suited to native-2-bit hardware.Entwine: Coordinating Tiled Computation and Fine-Grained Communication across GPUs
针对GPU计算与通信不匹配问题，提出Entwine方法协调分块计算与细粒度通信，以重叠计算和通信提升性能。
Authors: Kai Ma, Quanfeng Lv, Jingguo Ge, Bowei Dai, Kefan Ruan | Date: 2026-09-10
Link: http://arxiv.org/abs/2609.11562v1
Abstract
Modern high-performance GPU computations partition tensors into tiles to exploit data reuse and parallelism. Individual tile computations complete earlier than the full tensor computation, creating opportunities to overlap computation and communication. However, a mismatch between computation and communication progress can limit these opportunities. Communication stalls when no data is ready, and may lag when data arrives in bursts. Communication can also slow computation by consuming shared resources, offsetting the benefits of overlap. We present Entwine, which coordinates tile computation order, fine-grained communication, and SM resource allocation to minimize overall completion time. Entwine reorders tile computation to produce data for communication at a more regular pace. Entwine couples this schedule with fine-grained SM-based communication to process tile results with low latency and low overhead. Since the communication kernel also consumes SM resources, Entwine coordinates their allocation to balance communication progress against computation slowdown. Across representative tensor-parallel LLM workloads, Entwine achieves a geomean speedup of 1.232x (up to 1.433x) over cuBLAS+NCCL, and outperforms state-of-the-art overlap baselines by 3.1-9.8% in geomean. We will open-source our implementation upon publication.PATTON: Enabling Commodity PIM for Production LLM Serving
针对PIM在LLM服务中支持KV缓存生命周期管理的挑战，提出PATTON方案，解决内存分配与命令生成的冲突。
Authors: Hangyeol Kim, Sanghyun Lee, Teokkyu Suh, Joo-Young Kim | Date: 2026-09-10
Link: http://arxiv.org/abs/2609.11392v1
Abstract
Processing-in-Memory (PIM) is promising for accelerating memory-bound decode attention, but attention acceleration alone is insufficient for production LLM serving, where engines dynamically allocate, populate, share, cache, and reclaim logical KV cache blocks. Supporting this lifecycle on commodity PIM requires efficient physical memory allocation, block-to-address mapping, and command generation. For the Value cache, these requirements create a fundamental conflict among GEMV efficiency, single-token write efficiency, and memory capacity: GEMV-optimized layouts scatter newly generated Value vectors across rows, making writes costly, while finer-grained memory sharing improves capacity utilization but fragments GEMV reductions. We present PATTON, a PIM runtime that integrates production LLM serving engines with commodity PIM. PATTON introduces hierarchical granule allocation: block-sized Key and Value granules map one-to-one to logical token blocks, fixing their physical placements and commands, while coarser granules group blocks for efficient GEMV execution and memory utilization. A Commit Zone stages partial Value blocks for efficient single-token writes before committing them to GEMV-optimized locations. PATTON tracks these placements to generate KV cache writes and QK-transpose/SV commands. Across attention execution and runtime-induced prefill recomputation, PATTON achieves an average 1.95x speedup and 4.83x higher energy efficiency over evaluated baselines, requires no PIM processing-unit modifications, and maintains a KV cache hit rate comparable to the native GPU KV cache in vLLM.Memory Compression for High-Fanout Agent Sandboxes
针对高扇出智能体沙箱内存瓶颈问题，提出利用模板和跨沙箱冗余的内存压缩方法，解决传统压缩不匹配的问题。
Authors: Mengming Li, Ceyu XU, Qijun Zhang, Jiangnan Yu, Xiangfeng Sun, Haohui Mai, Zhiyao Xie | Date: 2026-09-10
Link: http://arxiv.org/abs/2609.11294v1
Abstract
High-fanout agent workloads create a growing memory bottleneck because a single task may spawn many concurrent sandbox sessions. Yet these sandboxes are far from independent: they originate from a shared template and execute related trajectories, exposing substantial template-relative and cross-sandbox memory redundancy. Conventional memory compression is poorly matched to this setting in three fundamental dimensions: how to compress, because they fail to exploit similarity across non-identical sandbox pages; what to compress, because they control page-fault overhead through conservative page selection; and when to compress, because compression is either triggered by memory pressure or performed without awareness of agent execution phases. We present AgentZip, the first memory compression system designed specifically for AI-agent sandboxes. AgentZip introduces compression mechanisms that exploit both the template-relative and cross-sandbox redundancy. It broadens the compression scope to any page with a profitable representation and shifts overhead control from compression-time page selection to restore-time prefetching. It further aligns expensive compression with LLM waiting periods to avoid interfering with foreground tool execution. Across LLM training and inference workloads, AgentZip reduces sandbox-owned memory by up to 8.7x, compared with 2.1x for the Linux configuration. Restore prefetching and agent-execution-aware scheduling reduce the slowdown of aggressive compression from as high as 3.1x to 1.40x while retaining nearly all of its memory-saving benefit.REVA: Reusable Evidence View Aggregation for Context-Efficient RAG Serving
针对RAG服务中上下文过长导致延迟和内存开销问题，提出可复用证据视图聚合方法REVA，实现上下文高效压缩。
Authors: Tuan Nguyen, Qiran Hu, Banruo Liu, Khoa D. Doan, Kok-Seng Wong, Fan Lai | Date: 2026-09-10
Link: http://arxiv.org/abs/2609.11209v1
Abstract
Retrieval-augmented generation (RAG) improves knowledge-intensive large language model (LLM) applications by conditioning generation on retrieved documents, but longer contexts increase latency, key-value (KV) cache memory, and token cost. Post-retrieval compression can reduce this cost, yet existing compressors often operate independently for each query, rely on auxiliary models or rewriting, and introduce online overhead that can offset the benefit of shorter prompts. We revisit RAG compression from a data-mining perspective by aggregating historical query--document--model interactions into reusable evidence views. We first show that modern compressors have unstable gains over simple truncation and can add substantial inference-time latency. We then propose Reusable Evidence View Aggregation (REVA), a framework that mines the target generator's historical attention traces into a document-keyed, budget-agnostic score store. REVA maps token-level attention to readable word units, aggregates importance across repeated document accesses, and renders budget-specific plain-text views that preserve document order and the standard RAG interface. Across four representative benchmarks and modern LLMs, REVA improves generation quality by 1.0--5.8 points over existing advances, while reducing compression overhead by a factor of 5.3 to 15.6, adding less than 40 ms of latency.FlexComp: One Model for Every Ratio in Context Compression
针对软上下文压缩中压缩比固定问题，提出FlexComp框架，通过马特廖什卡训练解耦压缩比，实现单一模型支持多种比率。
Authors: Kaiyan Zhao, Zhongtao Miao, Akiko Aizawa, Yoshimasa Tsuruoka | Date: 2026-09-10
Link: http://arxiv.org/abs/2609.11192v1
Abstract
Soft context compression condenses a context into a few memory tokens that a frozen LLM consumes in place of the raw text, but existing compressors fix the compression ratio at training and inference: each deployed ratio requires a separately trained model, and the chosen ratio is applied uniformly to all inputs, whose actual needs vary drastically. We propose FlexComp, a method-agnostic framework that decouples the ratio from both training and deployment: Matryoshka-style training samples the memory budget $K$ per instance, turning one model into an any-ratio compressor, and the budget is then chosen per input by: (1) confidence-based cascade routing or (2) a lightweight learned $K$ predictor. Across ICAE, 500xCompressor, and SAC on MRQA, a single FlexComp model matches separately trained fixed-ratio specialists with minimal degradation. Cascade routing preserves over 98% of the mildest ratio's accuracy at up to 266x average compression; the $K$ predictor, in a single compression-decoding pass, reaches 158-236x within 0.7 F1 of the mildest ratio. At serving-scale batch sizes, the $K$ predictor cuts context KV cache by 50% and improves decoding throughput by 47%.LILA: Calibration-Free Structured Pruning of Large Language Models via Latent Spectral Geometry
针对LLM结构化剪枝需校准数据问题，提出LILA方法，通过潜在谱几何分析神经元重要性，实现无需训练的剪枝。
Authors: Sankar Behera, Dhruv Singh, Anshika Agnihotri, Raj Kumar Choudhary, Satyadev Ahlawat, Yamuna Prasad | Date: 2026-09-10
Link: http://arxiv.org/abs/2609.11163v1
Abstract
Structured pruning of large language models (LLMs) offers hardware-efficient compression, yet existing methods require calibration data, gradient computation, or large auxiliary policy networks at pruning time. LILA (\emph{Latent-Informed Layer Analysis}) scores neuron importance via the Kolmogorov--Smirnov (KS) distance between empirical singular value distributions of the full and neuron-ablated feed-forward network (FFN) weight matrix, providing a closed-form spectral rule requiring no training, calibration data, or auxiliary network. Without any fine-tuning, LILA surpasses PruneNet (45M-parameter RL policy) by 1.57~pp in zero-shot accuracy on LLaMA-2-7B at 25\% sparsity, and outperforms WikiText-2-calibrated SliceGPT by up to 6.0~pp across all sparsity levels, while preserving the original architecture. After one epoch of LoRA recovery fine-tuning, LILA achieves highly competitive performance, matching the heavily calibrated SliceGPT baseline to within a 0.48~pp margin across LLaMA-2-7B and Phi-2, despite using zero calibration data. A Neural Tangent Kernel analysis confirms a 22$\times$ reduction in functional distortion versus random pruning, providing theoretical grounding for the spectral importance criterion. Finally, extending LILA to dynamically allocate sparsity budgets via KS-scores yields state-of-the-art generative preservation at moderate compression, while uncovering fundamental single-layer architectural bottlenecks at higher compression regimes.Phase-Decoupled, Model-Calibrated Power Control for Disaggregated LLM Serving
针对分离式LLM服务中功耗控制问题，提出相位解耦、模型校准的功率控制方法，优化吞吐量与延迟的权衡。
Authors: Jae Gon Kim, Donghoon Yoo, Hanyul Ryu, Sungho Ha, Juyeon Lee, Soojung Ryu | Date: 2026-09-10
Link: http://arxiv.org/abs/2609.11133v1
Abstract
Datacenter GPU power is the binding constraint on LLM serving capacity, and production serving has shifted to prefill/decode (PD) disaggregation. Deploying NVIDIA's Max-Q inference profile on a disaggregated B200 system, we found its realized gain modest (+8.6% tokens/J), model-dependent, and carrying a mean end-to-end latency cost (+5.2%) that throughput-only evaluation does not surface; the profile also applies one setting to prefill and decode GPUs that operate in opposite hardware regimes. We hypothesize that the optimal power setting is a property of the deployed (model, quantization, engine, hardware) combination rather than of the GPU class, that each lane warrants its own profile, and that converting SLO headroom into energy safely requires latency-gated calibration under a runtime SLO guard rather than a fixed recipe. We present a phase-decoupled, model-calibrated controller: the prefill lane runs under an SM-clock window whose floor is a latency guarantee by construction, and the decode lane under a power cap placed by automatic calibration just above a measured throughput/latency cliff. Because a disaggregated decode lane draws flat, memory-bound power, the cap binds continuously, the reactive-overshoot weakness that led POLCA to reject capping is absent, and the GPU's own power manager retains throughput under the cap. On an 8x B200 node serving Qwen3-Coder-480B (FP8) under agentic load, our balanced mode delivers +20.4% tokens/J at +3.5% mean e2e versus +8.6% at +5.2% for Max-Q, a Pareto improvement on both axes. On Qwen3-235B-A22B (NVFP4) every operating mode meets the ITL-p99 SLO in every repetition; both vendor profiles miss it. A decode-actuator A/B shows the calibrated cap beats static clock locks, and a three-day sustained run saves 32.3% of a lane pair's electricity. Both models are MoE; a dense model recovers roughly 5x less, so we scope our claims to MoE serving.EMMI: Edge Multi-Modal Intelligence for Communication-Efficient MLLM Inference via Fused Representation Compression
针对边缘多模态LLM推理的通信开销问题，提出融合表示压缩方法EMMI，降低计算、内存和通信需求。
Authors: Motahare Mounesan, Irfan Khan | Date: 2026-09-10
Link: http://arxiv.org/abs/2609.11058v1
Abstract
Recent advances in multimodal large language mod- els (MLLMs) have opened new opportunities for edge intelligence by enabling reasoning across heterogeneous sensor modalities, such as vision, text, and telemetry data. However, deploying these capabilities on resource-constrained edge platforms remains challenging due to the substantial computational, memory, and communication demands of modern MLLMs. Rather than transmitting raw sensor observations or partitioning neural networks at intermediate layers, Edge Multi-Modal Intelligence (EMMI) communicates a compact representation between edge devices and server resources, enabling communication-efficient edge MLLM inference. To achieve this, EMMI performs modality-specific encoding, cross-modal representation fusion, and learned compression at the edge, transmitting only a compact latent representation to server-side resources for high-capacity MLLM reasoning. This representation-centric design reduces communication overhead, preserves local data privacy, and provides a fixed-size interface between heterogeneous edge devices and server-side MLLMs. Evaluation on a representative multimodal benchmark demonstrates that EMMI can reduce the communication payload by 32x while maintaining comparable downstream accuracy, resulting in up to a 3.4x reduction in estimated end-to-end inference latency under bandwidth-constrained edge conditions.K/V-Cache Interventions Dissociate Representation Alignment from Persona Expression in Decoder-Only Language Models
研究K/V缓存干预对解码器语言模型角色控制的影响，发现表示对齐与行为表达存在解耦现象。
Authors: Yu Sun, Mengyin Lu, Cong Feng, Guangming Lu, Huimin Han | Date: 2026-09-10
Link: http://arxiv.org/abs/2609.11020v1
Abstract
We study K/V-cache interventions -- transplanting a target-conditioned K/V trajectory into a source-persona generation -- as a structured surface for persona control in decoder-only language models. Across 13 intervention configurations applied to Llama-3.1-8B for a fixed source-to-target persona pair, we report two consistent dissociations between representation-level alignment and behavioral expression, plus a common failure under position perturbations. First, all layer-band K/V replacements (early, mid, late) achieve strong local V-space alignment (V-gap 0.91, 0.89, 0.84), but only mid-layer replacement (layers 9-20) combines substantial target-marker expression with preserved lexical diversity. Second, full and mid-layer replacement induce comparable alignment (V-gap 0.94 vs. 0.89) yet produce different lexical-diversity profiles (TTR 0.65 vs. 0.77). Third, position perturbations (lag and shuffle) apply distinct operations yet uniformly suppress target-persona expression -- a common behavioral failure rather than a strict dissociation. Representation-level similarity metrics alone are thus not sufficient predictors of downstream persona expression in the regimes we study; the K/V cache emerges as a controllable but structurally constrained intervention surface. Because the transplanted trajectory carries the target's own generated token history, we characterize the intervention as trajectory-level transplantation rather than isolated persona-representation injection; a same-token-sequence control, decoding an identical token sequence under source vs. target conditioning, reproduces the sign and layer localization of the L28 representational shift, indicating the shift is not explained solely by imported token history. These findings characterize representation-behavior dissociation in a high-signal setting rather than establishing universality across models or persona pairs.Decoupling Readiness from Release for Tail-Aware Scheduling of Agentic LLM Workflows
针对智能体LLM工作流尾延迟问题，提出解耦就绪与释放的调度策略，通过尾风险感知调度降低尾延迟。
Authors: Bochao Feng, Jianjiang Li, Haojie Wang, Lin Qiao, Yinghui Li, Yukun Yan, Jidong Zhai | Date: 2026-09-10
Link: http://arxiv.org/abs/2609.10964v1
Abstract
Agentic LLM workflows consist of sequences of model turns interleaved with tool interactions, so their end-to-end completion time depends not only on inference speed but also on when ready turns are released. Most runtimes release each turn immediately upon readiness. Under contention, this eager release policy can accumulate released but unfinished work; once submitted, those turns can no longer be reordered by the workflow-level policy, increasing tail latency. We present a tail-risk-aware turn release scheduling method that jointly decides which ready turn to release next and how much released but unfinished work to maintain. The method uses a mean--Conditional Value-at-Risk (CVaR) objective to capture the evolving tail risk of unfinished workflows, incorporates online estimates of turn work when prioritizing ready turns, and adapts the released work budget to observed queue pressure. We evaluate the method using real agent execution traces from software engineering tasks across multiple LLMs and workflow arrival rates. The method performs comparably to eager release under light load and substantially reduces the P95 of workflow flow time under contention, achieving up to a \(3.50\times\) speedup.REACH: Controller-Managed Long-Span ECC for HBM AI Inference
针对HBM AI推理中纠错码开销问题，提出REACH控制器管理长跨度ECC，利用读主导特性降低解码成本。
Authors: Rui Xie, Yunhua Fang, Asad Ul Haq, Linsen Ma, Sanchari Sen, Swagath Venkataramani, Liu Liu, Tong Zhang | Date: 2026-09-09
Link: http://arxiv.org/abs/2609.10861v1
Abstract
High-Bandwidth Memory (HBM) cost motivates stronger controller protection that can support a wider range of device error rates. Long-span error-correcting codes provide stronger protection at a comparable code rate, but a direct implementation couples small accesses to span-wide state and requires costly decoding at HBM bandwidth. Read-dominated LLM decode offers a favorable setting: sequential reads support span aggregation, while sparse writes limit parity-update traffic. This paper presents REACH, a controller microarchitecture that uses established inner codes to correct common errors and identify unresolved chunks, reserving a long outer code for known-erasure repair. Differential parity bounds write traffic, and a co-designed endpoint preserves 32\,B transactions without an extra data burst. Ramulator2 sustains 1.88\,TB/s of application traffic at the highest error stress, while separate full-interface sizing supports a 2.69\,TB/s application target using ASAP7-synthesized kernels. At this analytical target, REACH's nominal composition uses 55.8\% less controller area and 57.7\% less modeled power than the evaluated mean-work direct-long design, showing the benefit of reserving long-span recovery for exceptional requests.Composable CXL Memory as a Kubernetes-Native Shared Memory for LLM Serving
针对LLM服务跨节点KV缓存复用问题，提出基于Kubernetes的CXL内存共享方案，实现可组合内存作为集群资源。
Authors: Hongjian Fan, Kevin Zhang, David Habinsky, Sean Dykstra | Date: 2026-09-09
Link: http://arxiv.org/abs/2609.10790v1
Abstract
We present a Kubernetes Dynamic Resource Allocation (DRA) driver that makes composable CXL memory a schedulable cluster resource, and evaluate the resulting shared-memory tier for cross-node KV-cache reuse in LLM serving. The driver composes CXL regions on demand, materializes them as DAX devices on each participating host, and injects them into pods under a single Container Device Interface (CDI) name so that pods on different nodes access the same physical region. A shared-memory connector for vLLM/llm-d uses that region as a KV-cache tier with a slot directory embedded inside the shared medium, which eliminates the need for an external metadata service. On a two-node cluster with a 512\,GiB CXL appliance and Qwen2.5-7B-Instruct, cross-node prefix reuse reduces TTFT by 5.5$\times$--36.6$\times$ at an external hit rate of 95.4--99.5\,\%, while node-local tiers (GPU prefix caching, CPU-DRAM offload) fall back to full recompute. The sharing gap, defined as the latency ratio between cross-node and same-node reuse, is 1--4\%, indicating that cross-node reuse incurs little additional latency relative to same-node reuse on our testbed. Both replicas run full engines; the study demonstrates memory disaggregation rather than prefill/decode disaggregation. We report this as a feasibility study rather than a performance evaluation.Why Is Video Still So Expensive? A Survey of Inference-Efficiency Mechanisms in Video and Audiovisual LLMs
综述视频和视听LLM的推理效率机制，分析其计算和内存成本随帧数和上下文长度增长的挑战。
Authors: Killian Steunou, Yannis Tevissen, Mounîm A. El Yacoubi | Date: 2026-09-09
Link: http://arxiv.org/abs/2609.10355v1
Abstract
Video understanding has rapidly evolved toward video large language models (VideoLLMs): systems that couple video representations with pretrained large language models and condition generation on a textual prompt. Their strong performance on captioning, question answering, retrieval and temporal grounding comes at a computation and memory cost that grows with frame count and context length, limiting deployment in real-time, mobile and resource-constrained settings. This survey covers inference-efficiency mechanisms for visual and audiovisual VideoLLMs that report concrete reductions in parameter count, FLOPs per input, latency, memory, or visual and audio token count. We analyze bottlenecks across frame sampling, modality encoding, connector-level token reduction, and LLM prefilling and decoding. We organize methods by the pipeline stage at which they act, covering VideoLLMs developed since late 2022 together with earlier frame-sampling and vision-encoder mechanisms that remain components of current pipelines. We assemble literature-reported accuracy--cost comparisons under shared host models and input protocols wherever available, distinguish them from heterogeneous cross-paper evidence, and identify gaps in audiovisual efficiency and standardized evaluation. We maintain a repository at https://github.com/momentslab/awesome-efficient-videollm.Beyond One-Size-Fits-All: Sample-Adaptive Strategy Routing for Vision Token Pruning in MLLMs
针对MLLM视觉token剪枝策略单一问题，提出样本自适应路由方法，利用策略互补性提升剪枝效果。
Authors: Haiji Liang, Pengfei Zhou, Zhenglin Wan, Wei Wang, Yang You, Wangbo Zhao | Date: 2026-09-09
Link: http://arxiv.org/abs/2609.10346v1
Abstract
Multimodal large language models (MLLMs) process hundreds or thousands of visual tokens per image, incurring prohibitive inference costs. While existing vision token pruning methods mitigate this overhead, they implicitly assume that a single fixed pruning strategy can be applied uniformly across all inputs. Our analysis further reveals that ranking pruning methods by average benchmark accuracy conceals substantial sample-wise complementarity: although the average-best strategy excels overall, alternative strategies prove superior on a significant fraction of individual samples. To harness this diversity, we propose VIP-Router, a lightweight VIsion Pruning Router that adaptively selects the pruning strategy predicted to be best suited to each input at a specified pruning level. Conditioned on low-cost visual and textual features, VIP-Router identifies the most suitable candidate strategy while retaining full-token inference as an option when pruning is predicted to be unfavorable. Evaluated on a curated suite of pruning-sensitive visual perception benchmarks, VTC-Bench Group A, VIP-Router consistently outperforms the best fixed strategy baseline across all reduction ratios, achieving a 26.9% relative improvement in average accuracy, and a 22.0% relative increase in average utility after accounting for realized token cost. Crucially, VIP-Router operates in a plug-and-play manner without modifying underlying pruning algorithms or model weights, introducing trainable parameters equivalent to merely 0.017\% of the backbone. Furthermore, VIP-Router proves effective across various MLLM backbones and yields consistent gains on unseen benchmarks, highlighting the potential of sample adaptive routing for visual token pruning.On-Policy Distillation for Vision-Language Model Adaptation, an Effective Paradigm on Low-Quality Multimodal Data
针对低质量多模态数据下视觉语言模型蒸馏目标不可靠问题，提出基于策略的动态蒸馏目标构建方法。
Authors: Hongyuan Zhang, Xianda Guo, Yanlun Peng, Qianlong Yang, Yubin Guo, Pinhan Fu, Mulin Chen, Xiaozhen Qiao, Ping Luo | Date: 2026-09-09
Link: http://arxiv.org/abs/2609.10321v1
Abstract
Knowledge distillation offers an efficient route to transfer a task-adapted vision-language teacher to a compact student. The training target in current vision-language distillation methods is typically constructed from the teacher prediction and applied uniformly to all training samples, making it unreliable under class and domain shifts. In this paper, we argue that distillation target construction should be treated as a dynamic training decision rather than a fixed recipe. To this end, we propose OnPoKD, an on-policy distillation framework for vision-language model adaptation. To the best of our knowledge, OnPoKD is the first framework that applies on-policy distillation to vision-language model adaptation by learning target construction as a policy decision. OnPoKD learns a lightweight controller that constructs sample-wise adaptive targets using reliability and disagreement cues from the teacher model, student model, and zero-shot prior. Instead of relying on a fixed teacher prediction, the controller dynamically balances teacher supervision, zero-shot prior guidance, and hard-label anchoring through bounded policy actions, allowing the distillation target to adapt to varying sample reliability and training stages. The policy controller is updated with validation feedback, encouraging target construction to optimize transferability rather than merely fitting the training distribution. Since the controller is only used during training, OnPoKD can be seamlessly integrated into existing vision-language distillation pipelines while preserving the original inference architecture and test-time cost. Extensive experiments on Base-to-novel generalization and Cross-dataset transfer benchmarks show that OnPoKD consistently improves over strong vision-language distillation baselines.RiLM: Parameter-Efficient Language Modeling via Geodesic Decoding
针对边缘部署小参数语言模型效率问题，提出RiLM方法，通过黎曼流形测地线解码移除输出层，提升参数效率。
Authors: Fang Li | Date: 2026-09-09
Link: http://arxiv.org/abs/2609.10305v1