The paper investigates whether training Mixture-of-Experts (MoE) routers can improve memory‑bandwidth locality on consumer GPUs. Using a new zero‑surgery telemetry tool, the authors measure that a large Qwen3‑235B model is bottlenecked by disk‑based expert access, and that an LRU cache can serve a majority of requests. They pre‑register experiments training 137 M‑parameter MoE models with locality‑aware losses, finding that while cache misses can drop up to 60 % (99 % static‑pin hit rate), every configuration fails to meet a strict 1 % perplexity threshold, indicating a tight coupling between cache efficiency and model quality.
By Shriniwas Ramesh Suram
The paper introduces two algorithms—CUCB‑SC‑MT and CLCB‑LS‑MT—for multi‑turn semantic caching in large language model serving. These methods learn a "reuse risk" metric from judge feedback to decide which conversation states to cache, balancing avoided generation cost against the risk of unacceptable reuse. Experiments on math‑tutoring and movie‑ticketing dialogues show that the online CLCB‑LS‑MT algorithm reduces generation cost by 35–41% and estimated latency by 19–35% compared to prior caching approaches.
By Baran Atalar, Selly Beyene, Xutong Liu, Carlee Joe-Wong
ValueDiff introduces a value‑geometric KV cache eviction strategy for large language models that suppress attention sinks. It ranks tokens by the L2 deviation of their value vectors from the cache mean, a score that aligns with minimal‑disturbance eviction under a max‑entropy assumption. Across several benchmarks—RULER, LongBench, and MATH‑500—ValueDiff consistently retains a higher proportion of useful tokens than prior methods, especially under tight cache budgets.
By Junyoung Park, Jungwook Choi, Mingu Lee
arXiv:2607. 04281v1 Announce Type: cross Abstract: Semantic caching reduces the latency and cost of retrieval-augmented generation (RAG) by serving cached answers to semantically similar queries, but most existing methods do not model the time-varying freshness of open-web evidence.
By Muhammad Mansoor, Tahir Ahmad, Yeo-Chan Yoon
RaReCache is a framework that allows a large target language model to decode accurately from a KV cache prefilled by a much smaller source model by selectively recomputing only critical token positions. It identifies these positions using a rank disagreement metric that scores tokens based on the energy of their mapped KV in output directions weakly supported by calibration data. Across two model families and five benchmarks, RaReCache achieves up to a 3.04× prefill speedup and significantly reduces prefill latency while maintaining 95–99% of target accuracy with a 30–40% recompute budget.
By Sreetama Sarkar, Saptarshi Mitra, Sitao Huang, Souvik Kundu, Peter A. Beerel
Serving a 235B-parameter Mixture-of-Experts (MoE) model on a single 8 GB GPU is bottlenecked not by compute but by memory bandwidth: decode must stream each token's active experts from whichever tier holds them, and on consumer hardware most experts sit on an SSD far slower than RAM. We quantify this bandwidth wall on Qwen3-235B (Q4_K_M, 134 GB): measured decode is 0.
The paper proposes invalidation contracts to manage cached recovery suggestions in LLM agents, attaching version stamps and cacheability hints to each suggestion so stale entries can be evicted without trial and error. The protocol separates realized savings into validity (protocol‑dependent) and compliance (planner‑dependent), showing that row‑level invalidation can significantly improve first‑try compliance and recover a substantial portion of token costs across multiple models, while table‑level invalidation can be detrimental. The study evaluates the approach across seven models, three serving paths, two domains, and about 9,400 episodes, demonstrating deterministic validity and high eviction precision.
By Michael Wu, Arquimedes Canedo
arXiv:2608. 07911v1 Announce Type: new Abstract: Mixture-of-Experts (MoE) models have outgrown accelerator memory, and offloading expert weights to host memory is now standard.
By Yu Zhang
The paper investigates how prefix caching, a default optimization in open‑source LLM serving stacks, affects reproducibility when combined with weight quantization. Experiments on an eighty‑episode multi‑turn agentic tool‑use workload show that enabling the cache causes the agent’s trajectory to change in 36.2 % of episodes at 16‑bit precision and 75.0 % at 4‑bit precision, while disabling the cache yields perfectly reproducible runs. The study identifies specific cache‑related settings that drive run‑to‑run divergence and demonstrates that cached serving is deterministic only when the cache state is preserved, which is not the case in typical deployments.
By Aditi Patodiya
arXiv:2608.07911v4 Announce Type: replace
Abstract: Mixture-of-Experts (MoE) models have outgrown accelerator memory, and offloading expert weights to host memory is now standard. This makes expert c...
By Yu Zhang
The paper introduces KVLearn, a learning‑based framework for managing key‑value (KV) cache retention in disaggregated large language model (LLM) serving systems. KVLearn comprises a Prefix Reuse Predictor that estimates reuse probability from prefix features, a Cost‑Aware Retention Score that translates this probability into keep/admit decisions based on recomputation, transfer, and storage costs, and an Adaptive Threshold Controller that tunes admission thresholds online. Integrated into a globally disaggregated serving topology, KVLearn achieves up to 56% faster time‑to‑first‑token and up to 53% reduction in inter‑node KV transfer compared to baseline caching strategies.
By Dong Liu, Yanxuan Yu, Eric Jiang, Shu Wang, Ying Nian Wu
Page-EntroKV is a new KV‑cache eviction framework designed for grouped‑query attention (GQA) in long‑context autoregressive language models. It pools token importance scores across query heads within each physical group using sink‑isolated Renyi‑2 entropy, then projects these pooled scores onto page frames and performs eviction at the hardware tuple level. The authors prove theoretical guarantees on union overhead, budget preservation, and needle‑retention, and demonstrate on a Qwen2.5‑1.5B‑Instruct pilot that Page‑EntroKV achieves exact union overhead control, eliminates sink masquerade, and retains 100% of critical tokens compared to mean‑pooling baselines.
By Inbasekaran S