Cache

KV cache memory has become a major deployment bottleneck for video generation and world models, which motivates low-bit quantization study for efficiency. Existing 2-bit KV cache quantization methods can achieve nearly lossless performance on video benchmarks such as VBench, however, we find that they still cause severe temporal flickering and visual degradation. Meanwhile, deeper investigates show that Key quantization produces smaller reconstruction errors than Value, but surprisingly leads to much larger output degradation. We trace this discrepancy to attention: small Key perturbations can change the attention logits, i.e., QK^\top, and shift the temporal-spatial tokens selected by Queries. These observations motivate us to explicitly preserve attention logits and temporal-spatial token selection during KV cache quantization to alleviate the visual degradation problem. To address this issue, we present QuantWM, a training-free and strictly causal 2-bit KV cache quantization framework. QuantWM introduces two complementary techniques to mitigate the attention shifts. Firstly, quantization-sensitivity-aware clustering (QSAC) jointly considers historical Query sensitivity and residual ranges to select INT2-friendly Key centroids, which reduces quantization errors in channels that are more critical to attention. In addition, principal-subspace attention compensation (PSAC) restores the remaining Key errors along the dominant Query subspace using low-rank projections, which provides a direct and efficient correction to stabilize attention logits. Extensive experiments on Causal-Forcing, LingBot-World-v2, HY-World 1.5, Matrix-Game-2 and Longcat-Video demonstrate that QuantWM significantly improves visual quality and temporal consistency, while outperforming existing methods across image and video quality metrics with up to 6.20x KV cache memory compression and limited additional overhead.

This paper builds a theoretical bridge between the economic notion of marginal utility and two machine-learning constructs, matrix factorization and the Key--Value cache of transformer language models. The singular value spectrum of a rating matrix is shown to be a diminishing marginal utility schedule for latent factors, the eigenvalue spectrum of the projected covariance operator to be the marginal utility schedule of a model's learned representation, and cache eviction and low-rank cache compression to be instances of constrained utility maximization under a memory budget. The three collapse into a single allocation rule: retain the top dimensions whose eigenvalue exceeds the shadow price of the binding constraint. The framework is applied to the automated extraction of structured information from geo-mining documents, where it motivates a multi-pass inference protocol, a layer-wise TIES model merging procedure, and a selection policy combining extraction quality, localization drift and energy, scalarized with a Conditional Value-at-Risk term on drift. Two empirical contributions are reported. An 11.2-million-parameter hierarchical classifier, trained in about five minutes on a single GPU, reaches 90.0 per cent level-1 accuracy on a held-out test set from a 973-document uranium-exploration corpus, against 92.0 per cent for a proprietary model on a fifty-document human audit of the same corpus, at a latency of 2.62 ms per card against approximately 2,000 ms for the API and at negligible cost. A diagnostic of uniform-density TIES merging exposes a reproducible degenerate mode in which the merged model returns token-identical outputs across five geographically distinct districts while declaring high confidence; re-executing the merge under layer-wise calibrated densities removes that signature on the diagnostic sample. The full-scale extraction benchmark, including LoRA fine-tuning, is reported as projected rather than measured and remains an empirical extension of this work.

The widespread adoption of long-horizon agents has made model workloads increasingly input-heavy. Although prior work has substantially reduced the cost of long-context computation, prefill remains computationally expensive, and large KV caches continue to strain HBM and SSD capacity and data-transfer bandwidth. Together, these compute, storage, and bandwidth demands constitute the primary bottleneck to further lowering deployment costs. To address this challenge, we introduce DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts (MoE) model with 552B backbone parameters and support for contexts of up to one million tokens. With its Causal Encoder-Decoder (CED) architecture, the model activates 16B parameters per token during decode but only 8B parameters during prefill, substantially improving cost efficiency for agentic workloads. To push the limits of KV cache compression, DeepSeek-V4.1-Flash combines cross-layer KV cache reuse in Compressed Sparse Attention 2 (CSA2) with FP4 KV caching. These designs reduce its global KV cache footprint (always in HBM) to 890 bytes per token, roughly 1/4 of the corresponding footprint of DeepSeek-V4-Flash. Further, through a dedicated deployment optimization known as SWA Bounded Replay, DeepSeek-V4.1-Flash reduces its persistent KV cache footprint (always on SSD or in host memory) to roughly 1/8 of that of DeepSeek-V4-Flash. Despite its much smaller KV cache footprint, the model delivers substantially better performance than the baseline. In addition, we streamline the DeepSeek-V4 architecture and introduce several efficient architectural extensions. We pretrain DeepSeek-V4.1-Flash on a multimodal corpus comprising 45T tokens and conduct comprehensive post-training, yielding strong performance across diverse text-based and multimodal agentic scenarios. Model checkpoints are available at https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash.

The KV cache has become a major bottleneck in deploying LLMs, as its memory footprint grows linearly with sequence length and batch size, imposing substantial pressure on both memory capacity and bandwidth. Among various KV cache compression techniques, quantization is particularly attractive due to its effectiveness and ease of deployment. However, most existing methods rely on fixed-width quantization, where a $b$ bit representation is inherently limited to $2^b$ quantization levels. As the bit width decreases, the number of available levels shrinks exponentially, leading to severe information loss and rapid performance degradation. We further observe that fixed-width quantization fails to exploit the highly non-uniform distribution of KV cache. After rotation and normalization, KV values approximately follow a normal distribution, with most values concentrated near the center and only a small fraction appearing in the tails. Nevertheless, fixed-width coding allocates the same number of bits to frequent and rare symbols. Entropy coding naturally exploits such non-uniformity by assigning shorter codewords to frequent symbols and longer ones to rare symbols, substantially reducing the average number of bits required for representation. However, its variable-length output is not suited to highly parallel attention kernels, where efficient dequantization and computation rely on regular memory layouts and fixed-stride accesses. To bridge this gap, we propose \textbf{D-Quant}, a flexible KV cache quantization framework that introduces a \textbf{drift} mechanism to convert entropy-coded representations of each token into fixed-size bitstreams, enabling regular memory access and parallel dequantization within attention kernels.

Autoregressive (AR) video generation degrades over long horizons due to an overlooked train-inference discrepancy we term KV eviction mismatch: models train on short clips where all context frames reside in the KV cache, but at inference, memory constraints force distant frames to be evicted from the KV cache - removing context the model was conditioned on. Rather than simulating eviction via context truncation - which discards temporal information the model still needs and degrades motion coherence - we keep the context but while progressively reducing the influence of distant frames, making their eventual eviction negligible. To guide this design, we introduce the positional response $R( Δ, \, t_{\text{denoise}})$, a perturbation-based sensitivity measure revealing that context influence decays steeply with temporal distance and varies systematically across denoising steps. Motivated by this analysis, we propose Recency Forcing, which applies a non-positive, timestep-dependent bias, termed Temporal Response Bias (TRB), on pre-softmax attention logits derived directly from $R$, closing the train-inference gap without modifying context length or training objectives. We further introduce Biased Attention Reparameterization (BAR), an exact reformulation that moves the bias outside the softmax, making TRB a standard FlashAttention call at zero overhead. Recency Forcing operates in both training-free mode and training-based mode. Experiments on VBench and VBench-Long demonstrate state-of-the-art long-horizon generation quality at no additional inference cost.

Repeated prompt prefixes are increasingly common in LLM serving workloads, appearing in system prompts, templated retrieval-augmented generation pipelines, agent frameworks, and multi-turn conversations. Modern inference runtimes such as vLLM and TensorRT-LLM provide mechanisms for reusing previously computed KV-cache state across requests, yet it remains unclear when prefix reuse materially improves serving performance on contemporary accelerators and when its benefits are limited by scheduling, cache granularity, concurrency, or memory pressure. This paper presents PrefixBench-H100, a reproducible benchmark and measurement framework for characterizing prefix reuse on a single NVIDIA H100. PrefixBench-H100 combines controlled synthetic traces with chat-style and retrieval-style workloads, and evaluates two widely used LLM serving runtimes under matched workload conditions. The benchmark varies shared-prefix length, suffix diversity, request arrival pattern, concurrency, output length, and cache configuration, while collecting time-to-first-token, inter-token latency, end-to-end latency, throughput, cache-hit statistics, GPU memory usage, and selected profiling traces. The goal of PrefixBench-H100 is not to introduce a new caching algorithm, but to expose the practical operating envelope of prefix reuse for H100-class LLM serving. The study identifies the regime where prefix reuse provides substantial first-token latency reductions and the regime where cache pressure erodes them, while showing that cache effectiveness itself is largely insensitive to concurrency and output length; the cross-runtime differences that remain arise above the cache, in the scheduling layer.

Autoregressive transformer decoding is constrained by irregular key-value (KV) cache movement on tiled accelerators. Prior compression and DRAM-placement systems still concentrate traffic on centralized memory paths that bottleneck long-context serving. We present MeshKV, a KV cache fabric that moves blocks as packetized flows over a lightweight NoC. It co-designs (i) TaKV affine striping to spread homes and cut hotspot load, (ii) Mare multicast with verified duplicate suppression, and (iii) Pad, which overlaps prefetch, tile multiply, and streaming softmax behind credit-aligned FIFOs. Together they convert bisection back-pressure into useful KV transfer. On our 8x8 FPGA implementation with LLaMA-2-7B and Mistral-7B at 8K-32K, MeshKV reduces interconnect traffic by up to 58%, improves KV bandwidth utilization by 2.1x, and delivers up to 1.9x multi-stream throughput.

KV-cache reuse can reduce inference cost in retrieval-augmented generation and agentic systems, but cached contexts may become stale when retrieved knowledge, working memory, or user state is edited. Under causal self-attention, even a local edit can affect downstream KV states. A full re-prefill reliably restores consistency but is costly, whereas refreshing only the edited span can leave downstream dependencies stale. We formulate in-place repair as budgeted recomputation and compare training-free position-selection policies on a factual RAG benchmark with matched direct and derived edits. Across three model families, all policies repair direct cases, but derived cases clearly separate them. At the primary budget, a contiguous edit-local window recovers at least 0.94 of the post-edit answer margin and substantially outperforms attention-based, KV-deviation, and structural selectors. Mechanistic analysis shows that position sets effective under clean-state transplantation can fail under actual recomputation because scattered positions inherit surrounding staleness. The edit-local advantage also depends on adjacency and largely disappears when the answer-bearing text moves downstream. Because answer-relevant edits almost always corrupt model behavior, failure severity is difficult to predict, and repair is 13-21 times faster than full re-prefill, our results support unconditional edit-local repair when the dependent text remains adjacent to the edit.

LLM inference optimizations report speedups on different models, GPUs, prompts, and quality metrics, making them hard to compare or combine. We build a cost, quality, and latency Pareto atlas to identify the best configurations for different deployment constraints. Since exhaustive testing is impractical, we measure 54 configurations of Qwen2.5-7B-Instruct running on vLLM 0.12 across L4, A100, and H100 GPUs and use these anchors to calibrate a simulator. It reproduces measurements at anchored batch sizes, with cross campaign drift below 1.5 percent. A separate quality evaluation tests FP16, AWQ 4bit, FP8 weights, and FP8 KV cache on 200 GSM8K questions with five examples per prompt. Sparse attention is evaluated only in simulation. On the calibrated grid, 18 of 36 configurations reach the Pareto frontier. Combined methods reach it more often than individual methods, with 9 of 15 combinations versus 9 of 21 single methods. Quality testing changes the winners. AWQ 4bit reduces per token latency to 0.34 times baseline on L4 but loses 5.9 percent of strict GSM8K accuracy, narrowly missing the 95 percent quality floor within sampling uncertainty. Flexible answer extraction matches FP16 accuracy, suggesting the loss comes from formatting rather than arithmetic. FP8 weights retain 99.4 percent of baseline accuracy at 0.61 to 0.65 times baseline latency across all three GPUs and appear in three of four regime winners. A naive FP8 KV cache maintains normal throughput but answers none of the 200 questions correctly, showing why speed alone is insufficient. Under two prompt designs, n gram speculative decoding measures at 0.90 to 0.98 times baseline and adds no benefit on this stack. The best choice depends on the constraint and GPU: H100 wins for tight latency, while A100 wins for throughput and low cost at 0.106 dollars per million tokens.

When agentic sessions run to a million tokens with many sessions resident at once, the KV cache and the index that ranks it live in host memory, and the scan that ranks all n keys for a top-k step becomes the traffic that bounds decoding. We present Fathom, a key scan in which each query decides how many bits of each key channel to read. The 4-bit K cache is stored channel-major as bit planes, so a prefix of t planes is exactly the channel's t-bit quantizer, and the query spends its bit budget by reverse water-filling over the variance-weighted importance of its channels. At one million tokens on Qwen3-8B a decode step is 1.67x faster in GPU time than with the 136-bit scans of Double Sparsity, Loki and SparQ r=32, and in the same GPU time as SparQ's 68-bit read (r=16) Fathom reads 18% fewer bytes with lower attention error on six of seven model and context settings. On RULER-style tasks every per-token scan matches exact top-k decoding, and on real coding-agent sessions Fathom reaches the step agreement of the most accurate 136-bit scan at 92 bits. The store is the 4-bit K copy a quantized serving stack already holds, and the method is not faster when the index is resident in GPU memory.

A common small-model deployment runs one shared backbone with several LoRA specialists that answer over the same context. Serving them naively re-prefills that shared context once per specialist. We study a narrow, practical question: for already-trained standard LoRA adapters -- not adapters retrained for cache compatibility -- how much task quality is preserved if the backbone's prefill KV cache is computed once and reused across specialists, and what does that buy in serving cost? On a Qwen3-1.7B backbone with two adapters (extractive QA on HotpotQA, arithmetic reasoning on GSM8K), we sweep the boundary at which the specialist takes over from the reused base cache and measure paired quality differences and serving cost. Full-prefix reuse had the lowest prefill cost and a small quality difference on held-out GSM8K (Delta = -4.6 EM at a 160-token budget; -3.0 at 320 tokens; -0.8 under a second training seed -- all favoring native, only the first excluding zero, and the magnitude not consistent). Partial recomputation provided no demonstrated advantage. Neither quality equivalence nor a general boundary-selection rule is established. We also report a closed-form ridge KV translator that did not beat direct reuse, and specialist-dependence contrasts whose intervals all include zero. The measured serving benefit is warm-cache time-to-first-token, which grows with context (~16x at 8K); two-branch peak memory was only 12% lower and, on inspection, the prefix was never physically shared across branches -- this implementation reuses KV values but copies their storage, so shared-cache memory savings are not achieved.

Diffusion large language models (dLLMs) offer a promising parallel decoding paradigm as an alternative to autoregressive generation through iterative unmasking. However, dLLMs typically require many steps before token confidence reaches the decoding threshold, resulting in inefficient inference even with block-wise KV caching. To accelerate dLLM inference, we for the first time propose an "early-bird (EB)" decoding framework, motivated by the observation that tokens with similarly low entropy tend to cluster and can be jointly decoded earlier, before reaching the confidence threshold. In particular, our EB-Decode framework integrates two key enablers: (1) a learnable network that adaptively groups tokens with similar uncertainty into variable-length blocks, rather than relying on fixed block sizes; (2) a position-aware sampler that learns to unmask tokens in parallel using fewer decoding steps within predicted variable-length blocks. Both components are developed without modifying pretrained dLLM weights and can therefore be directly deployed as plug-ins during serving, with negligible training and inference overhead. Extensive experiments across three models and four benchmarks consistently validate our observation and the effectiveness of EB-Decode, achieving 3.53-18.76$\times$ higher throughput than the vanilla decoding method and up to 1.58$\times$ higher throughput over the strongest baseline, Fast-dLLM, with comparable accuracy.

Wireless edge caching networks typically consist of many independent Base Stations (BSs), each facing its own request rate and content popularity profile. Training a Reinforcement Learning (RL) caching agent from scratch at every BS forces each agent to relearn, through slow trial and error, a decision problem that is structurally identical across the network. Meta-reinforcement learning removes this redundancy by learning a shared initialization that adapts to any BS in a few local updates; however, meta-training itself becomes the bottleneck at scale: the meta-gradient must be estimated from a small subset of BSs at each meta-iteration, and sampling this subset uniformly at random yields a high-variance estimate, an issue existing meta-RL caching frameworks leave unaddressed. This paper proposes a meta-reinforcement learning framework for caching across independent, non-overlapping BSs that directly targets this bottleneck. Each BS runs a local Proximal Policy Optimization (PPO) agent, formulated as a Semi-Markov Decision Process (SMDP) over content popularity, size, lifetime, and importance, while a shared meta-policy is learned via a Model-Agnostic Meta-Learning (MAML)-style loop. To scale meta-training and accelerate convergence, we introduce gradient-based clustering, which groups BSs by local gradient similarity and draws from every cluster, in proportion to its size, at each meta-iteration. We prove, via an Analysis of Variance (ANOVA)-style decomposition of gradient variance, that this strategy yields a strictly lower-variance meta-gradient estimator than uniform random sampling under BS heterogeneity.

Large Language Models (LLMs) have shown impressive capabilities across a range of natural language processing tasks, and LLM inference has emerged as a critical workload for enabling downstream applications. The demands of serving LLM inference are becoming increasingly challenging as requests shift toward longer sequences and heavier inference, driven by retrieval-augmented generation, inference-time compute scaling, and long-context applications. Additionally, these challenges are compounded by hardware trends, as memory capacity and communication bandwidth are not scaling as fast as increases in workload complexity. Compute-in-Flash is a promising solution to address memory bandwidth limitations by moving computation close to memory, and to exploit the large capacity of SSD technologies. However, it is challenging to deploy LLMs on these systems as they lack support for high-precision floating point operations and have limited write endurance. In our work, we aim to address these challenges by designing inference algorithms to enable LLM inference on Flash compute-in-memory devices. We present an end-to-end integer-only quantization approach to eliminate expensive floating-point computations. To address the limited write endurance, we design a dictionary-based KV cache compression strategy based on sparse dictionary coding that represents each KV vector as a linear combination of static dictionary vectors. These algorithmic improvements enable us to exploit the benefits of Compute-in-Flash for both model weights and KV cache, and to minimize expensive data transfer operations. Across Llama-3.1-8B and Qwen-2.5-7B, our combined method exhibits limited accuracy degradation while reducing dynamic KV cache traffic by 15$\times$.

Looped Transformers have recently demonstrated strong performance in both reasoning and language tasks by reusing a shared set of parameters across multiple iterations, achieving parameter efficiency without sacrificing representational power. Besides, looped Transformers perform inference directly in the latent space (latent reasoning) to reduce the number of tokens consumed during inference, thereby achieving improved sample efficiency. However, these models typically apply a fixed recursion depth uniformly to every token, leading to suboptimal compute allocation and leaving significant efficiency gains on the table. In this work, we propose \textbf{dynamic token-choice routing} for looped transformers, enabling each token to adaptively determine its own number of loop iterations based on its hidden state. We use a dynamic router to decide whether a token should continue recursing or exit early, allowing simple tokens to bypass unnecessary computation while hard tokens receive deeper processing. To ensure that this adaptive mechanism does not compromise decoding efficiency, we further introduce recursion-wise KV caching, which maintains an independent key-value cache for each recursion loop. This design ensures that tokens at different depths only attend to their corresponding cached states, effectively eliminating redundant computations for exited tokens and enabling fast autoregressive decoding. Extensive experiments show that T-LoopFormer reaches the sota performance under the same parameters on PPL and 10 zero-shot reasoning tasks, even surpassing the base model at 24x FLOPs and our model could reach the lowest inference latency, which validate the effectiveness of token-choice router and recursion-wise KV cache. Code: https://github.com/YuMingQian1234/T-LoopFormer.

Shared host-memory caching can avoid repeated prefill when a request moves between inference replicas. Its usefulness depends on both correct state transfer and lost prefix locality. We study two single-GPU 27B vLLM replicas sharing a 256 GiB LMCache pool. After adopting an existing packed-page patch, we isolate a raw-pointer fallback that omits the dependency on the current CUDA stream. Controlled byte tests fail under an imposed delay and pass when the dependency is restored; the existing mixed allocator provides a working deployment path. Full-pool allocation checks and service regression complete the validation. A four-block OFF-ON-ON-OFF comparison contains 768 measured requests within two block pairs. Median cross-replica time to first content token falls from 31.715 to 0.605 seconds at 128k input and from 92.047 to 0.790 seconds at 256k. Six-turn synthetic sessions alternating replicas improve by approximately 35% and 45% at initial contexts of 32k and 128k, while fixed placement shows little benefit. This engineering case study identifies practical validation steps and the locality conditions in which shared caching pays off.

Diffusion language models decode tokens in parallel, but their bidirectional denoiser rules out the naive key--value (KV) cache behind fast autoregressive inference. Block diffusion restores caching by decoding block-by-block, and the block caches deployed on it so far are tied to attention: O(L)in memory and, if used as training-free retrofits, only an approximation of the model's computation. Both constraints can be overcome: sequence mixers that summarize finalized blocks into a reusable state support block caching, and the corresponding block-causal training objective makes the cache exact. We study this recipe at scale, pretraining three 3B block-diffusion denoisers (attention, mamba, and hybrid) on 300B tokens under one single-frontier objective and decoding all three through a single cached interface. Only the state-space cache is O(1) in sequence length: its memory and per-step latency stay constant at any context length, while an attention cache remains O(L). At 256k tokens (where attention has grown to 82GB and 29 ms/step), the Mamba cache delivers 4.3x lower latency, 11x less memory, and 2.6x higher single-stream throughput; and because that footprint is constant it scales with batch as well, reaching 14x the aggregate throughput, where attention cannot run beyond a single stream. The same linear-state bias lets the Mamba and hybrid backbones keep retrieving out to 8-16x their training length, whereas attention's retrieval collapses at 2x, at no measured quality cost.

Vision-language models retain a substantial amount of pixel-decodable visual content in their visual key-value cache. We show, in our setting, that this retention is task-inert: across our preregistered tests, how much a unit retains never positively tracks whether the computation that answers the question causally relies on it. We measure retention with a learned pixel-inversion decoder and causal use with single-super-patch KV ablation, the teacher-forced drop in gold-answer log-probability, and relate the two within images under a preregistered, sign-calibrated, held-out design. Retention is decoupled from attention and, in a well-powered null, from causal utilization. Utilization is not inert to every proxy: attention weakly but significantly tracks it, the only signal we find that does and the design's positive control. We characterize pixel-decodable retention as an informational axis of the visual KV cache, orthogonal to the functional one. How much task-inert content a cache holds differs by architecture in our model pair: the encoder-free model retains 2.7 times more than the encoder-based one. The engineering consequence is a controlled negative result. At super-patch granularity, deconfounded pixel-decodable retention ranks KV eviction no better than random; at token granularity it acquires only a weak inverse-importance signal at larger budgets, dominated at every budget by attention magnitude. In our setting, pixel-decodable reconstructability is not a competitive KV-compression signal at any granularity we test.

With the recent rise and adoption of tabular foundation models, optimizing their inference performance becomes an emerging field for efficiency research. While the models are architecturally similar to transformer-based large language models (LLMs), the size and serving patterns differ significantly. We show that the focus should be on the attention calculation and less on weight or KV cache quantization, which are more popular in LLMs. We develop a quantization strategy for queries, keys, and values to FP8 and use explicit FP8 matrix multiplication instructions to speed up the attention calculation. We find that it is crucial to align the quantization error in the test rows with the quantization error in the training rows, as otherwise the accuracy drops drastically. Our Triton kernel achieves a speedup up to 1.7x over regular 16-bit kernels, and we show that on TabPFN-v3 and TabICLv2 there is no relevant accuracy loss across TabArena and BeyondArena.

Text and lyrics specify broad musical characteristics and sung content but offer limited control over musical timing, melody, and reference-based style. We introduce DiffSynth-Music (https://modelscope.cn/models/DiffSynth-Studio/DiffSynth-Music), a framework that adds composable audio conditioning to a music synthesis backbone through layer-wise key-value injection. The three template models, Control, Prosody, and Reference, are initialized from the backbone diffusion transformer and trained with conditional flow matching. They support five control types: beats, vocals, accompaniment, prosody, and reference audio. A shared variational autoencoder maps conditioning waveforms into a common latent space, enabling their attention memories to be combined. With the template timestep fixed at the clean-data endpoint and other inputs held constant, each control cache is computed once and reused throughout sampling. Training pairs are derived from music recordings using beat extraction, source separation, vocal resynthesis, and reference-excerpt selection. Single-control evaluations on Mandarin and English songs demonstrate improved adherence across all five control types and better lyric fidelity under vocal conditioning relative to the backbone. Automatic music-quality and instruction-following scores remain broadly comparable to those of the evaluated base models, with metric-specific trade-offs. We release the three template models to support research and creative applications in controllable music generation.

Agentic serving can consume orders of magnitude more tokens than chatbot workloads, stressing both KV-cache capacity and decode-time bandwidth. Most KV-eviction methods score cached keys against representative queries drawn from the most recent tokens, assuming future attention resembles recent attention. We show that agentic generation violates this assumption: future queries form a mixture over think, act, tool, and others phases, and principal-angle analysis shows these components occupy measurably different query subspaces, so recency representatives systematically undervalue keys that upcoming phases will need. We propose AGENTKV, which maintains a small query buffer per phase and scores cached keys against their union. We further implement AGENTKV in a persistent multi-turn serving path that carries compressed KV state across turns and compacts retained KV pages online. Across two models, six task domains, and three KV budgets each, AGENTKV improves task score by 5.5 points on average over R-KV and 5.3 over Tri-attention. Relative to upstream full-KV SGLang, AGENTKV improves output-token throughput by up to 1.80x. Code: https://github.com/LiuTaowen-Tony/agentkv.

As Omni-modal large language models (Omni-LLMs) take in audio, video and text together, their KV cache memory cost grows. KV cache quantization is the de facto approach in text-only LLMs, but its application to Omni-LLMs remains unexplored. In this paper, we analyze how TurboQuant, a representative rotation-based KV cache quantization method, behaves on multimodal caches and identify two critical issues: temporal key drift and heterogeneous value geometry. To address these, we propose OmniKVQuant, a training-free framework that (i) sets the key quantization range over each short window of the input stream; and (ii) rotates values separately per modality. On Qwen2.5-Omni and Qwen3-Omni, OmniKVQuant enables 2-bit KV caches while substantially preserving performance across seven audio-visual benchmarks. We further provide a fused Triton decode kernel that unpacks the 2-bit cache during attention, so no dense FP16 cache is ever built. Code: https://github.com/kaistmm/OmniKVQuant

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.

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.

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.

GUI agents accumulate high-resolution screenshots as the trajectory unfolds, increasing inference latency and memory usage. Training-free visual token pruning can reduce this cost, but cache reuse introduces a fundamental constraint. Once tokens are discarded, the corresponding visual evidence cannot be recovered without re-encoding. Pruning therefore becomes an \textit{irreversible admission decision} that must remain useful for unknown future targets while preserving coverage of operable regions under tight budgets. To address these challenges, we propose \textbf{\method{}}, a training-free framework for \emph{\textbf{T}rajectory-\textbf{r}obust \textbf{A}dmission and \textbf{C}overage-aware \textbf{E}vidence ordering}. Specifically, we combine a query-independent layout-derived interaction prior with instruction relevance and feature novelty to rank visual evidence according to both potential future utility and diversity. Then, we reserve part of the budget for native visual tokens distributed across the screen, repairing missing spatial coverage without breaking the ordering. Together, these mechanisms produce a nested token order, allowing retained visual evidence to shrink monotonically across budgets while remaining reusable throughout the trajectory. Finally, our monotone KV contraction incrementally contracts retired frames into compact session state, avoiding repeated visual encoding or pruning. Extensive experiments across six GUI benchmarks and diverse models verify the effectiveness of our proposed \method{} under tight budgets. The source code will be released.

LLM serving systems already reuse KV caches, but only when the reused text sits at the very start of the prompt. Two growing workloads break this condition: a retrieval-augmented generation server assembles a different set of retrieved chunks for every query, and a multi-agent coordinator reads reports written by other agents. Reused inside a new prompt, a cache carries the wrong positions and never attended to the other sources. The cache may also have been written by a different checkpoint of the same model family, which changes the stored values. Repair methods for such caches have appeared in three separate communities, each measured on its own terms, and existing benchmarks test only exact-prefix reuse, where nothing is lost. KVShareArena benchmarks KV-cache reuse across prompt contexts and model checkpoints on retrieved chunks and agent reports. It scores every method by the fraction of the gap it recovers between no cache and full recomputation, and charges compute, memory, and per-request latency with the cache in hand, reporting the one-time cost of building a cache separately. We find that correcting positions, which needs no recomputation, is enough until a question needs several sources at once. There, only methods that pay, by re-encoding part of the cache or by training, recover half to two thirds of the gap; unrepaired caches can be worse than no cache. Cache-compression methods that are harmless on a single prompt fall significantly behind position correction on freshly written agent reports. These patterns hold across three model boards. When a different checkpoint wrote the cache, training-free methods are barely affected, while an adapter trained on one checkpoint's caches loses quality. Harness, frozen querysets, and cost accounting ship as a pip package with an automated submission workflow and a public leaderboard.

In Retrieval-Augmented Generation (RAG) systems, a large number of retrieved chunks are concatenated to form the input context so that users can receive high-quality responses based on external knowledge. As a result, the input context length increases substantially, leading to a larger prefill workload and, in turn, a longer time to first token (TTFT). While previous works that reuse precomputed key-value (KV) caches effectively reduce TTFT for long-context inputs, it remains unclear whether response quality is preserved when the input context becomes very long. In this paper, we propose a combined approach that (i) fine-tunes the model while taking KV cache concatenation into account and (ii) selectively recomputes a subset of the KV caches. By applying both techniques, we demonstrate improved accuracy for long-context inputs. Experiments on the RULER benchmark show that, for a 124k-token input, our method improves the RULER score by 9.7 point over the baseline that recomputes KV caches only. Moreover, TTFT is reduced by 80% compared with full attention.

The KV cache is a primary bottleneck for Transformer decoding: its memory footprint and cache-read traffic grow with sequence length. Grouped-query attention (GQA) reduces this cost by sharing key-value heads, but still stores both a key and a value at every step. We introduce Grouped Value Attention (GVA), which stores grouped values and reconstructs content keys with a learned linear map. At inference, the map can be absorbed into the query, eliminating the need to materialize content keys in the intended decode path. A small shared decoupled RoPE channel retains positional information through a separately cached positional key. For the configurations studied, this representation reduces persistent cache scalars by approximately 45-47% relative to matched GQA. At the 350M-parameter scale with 30B FineWeb-Edu tokens, the 16-dimensional positional variant reaches 44.18 average accuracy across five tasks, compared with 44.36 for GQA and 43.88 for MLA. These results demonstrate near-GQA benchmark accuracy with a more compact cache representation. To translate this compact representation into faster autoregressive inference, we have developed custom decoding kernels and are currently evaluating their end-to-end inference performance with an open-source release planned soon.

KV cache is evolving from a serving optimization into an external memory substrate for long-term LLM agents. In a shared multi-user deployment, however, reusable KV blocks introduce a missing access-control question: semantic relevance alone cannot determine whether a memory block is authorized for the current physical user. We propose Bio-MemArt, a biometric-aware KV-cache memory framework for multi-user LLM agents. Bio-MemArt attaches a normalized biometric template to each stored KV memory block, filters the shared memory pool with the current user's biometric probe, and then runs the original MemArt retrieval and KV reuse pipeline only inside the authorized candidate pool. This design preserves latent-space retrieval, direct cache reuse, and decoupled position encoding while adding physical-user access control to shared KV memory. We evaluate Bio-MemArt under Owner and Non-owner query conditions on long-term dialogue QA with face and palmprint benchmarks. Across face benchmarks, the average owner and non-owner biometric success rates are 95.71% and 0.86%; across palmprint benchmarks, they are 97.60% and 2.00%. In the efficiency study, average prefill tokens drop from 18,781.96 under full-context prompting to 28.57 with Bio-MemArt, showing that biometric gating preserves the low-token operating regime of KV-cache memory.

Agent memory systems must discard stored information when their history exceeds a fixed token budget. Existing budget-accuracy frontiers quantify the resulting loss in accuracy, but do not distinguish irreversible losses caused by eviction from recoverable retrieval failures. We introduce the restore counterfactual, a per-question paired intervention that reinstates the question's gold evidence in the read-time context and reruns the same reader. Combining the change in correctness with whether the evidence was retained after eviction classifies each oracle-answerable error as recoverable, irreversible, or residual; in the residual case, the answer remains incorrect after restoration. We evaluate FIFO, random, redundancy-aware, and LLM-importance eviction on LongMemEval-S at three budgets and under two retrieval regimes, using GPT-4o-mini as the primary reader and judge and GPT-5.4-mini as a robustness reader. Under top-k retrieval at an 80k-token budget, the irreversible share among errors corrected by restoration is 0.67-0.73 for FIFO, random, and redundancy-aware eviction, compared with 0.60 for LLM-importance. At 8k tokens, it reaches 1.00 for all four policies. Recoverable errors occur under top-k retrieval at 80k tokens but are absent under forced-gold injection by construction, so budget-accuracy results are not directly comparable unless the retrieval regime is reported. An exploratory matched-accuracy analysis detects no difference in irreversible rate among accuracy-matched policy pairs at a resolution of 1.2-6 percentage points. The same analysis detects the deliberately destructive control. To our knowledge, this is the first per-item, per-question restore-counterfactual audit of eviction for external agent-memory stores on a standard conversational benchmark.

Key-value (KV) cache eviction is essential for scaling long-context inference in Large Language Models. However, existing policies predominantly rely on empirical heuristics, lacking a rigorous characterization of token utility under the inherently nonlinear softmax attention mechanism. In this work, we rethink KV cache eviction through the lens of local information geometry, modeling the attention process as a nonlinear Gaussian communication channel. By performing a first-order Taylor expansion of the attention mapping, we derive the Jacobian Information Capacity, a novel objective that explicitly captures query relevance, softmax sensitivity, and structural diversity. Guided by this theory, we introduce Jacap, a capacity-aware eviction method that utilizes softmax-aware importance weighting and statistical leverage scores for subset selection. Extensive experiments across diverse architectures and benchmarks demonstrate that \textsc{Jacap} delivers superior performance in most scenarios, particularly in high-compression regimes.

Key--value (KV) cache compression is an effective way to reduce the memory overhead of large language model (LLM) inference, particularly for long-context workloads. However, existing compression methods make different trade-offs among accuracy, inference latency, and peak KV cache memory utilization, making a single fixed configuration unsuitable across different prompts and resource constraints. We introduce MetaKV, an adaptive framework that selects a KV cache compression configuration for each input prompt based on user-specified latency and peak memory budgets. MetaKV uses lightweight prediction models to estimate the end-to-end latency, peak memory, and probability of a correct response for each candidate configuration, and selects the configuration that best satisfies the latency-memory constraints while preserving accuracy. We evaluate MetaKV across ten configurations from three representative KV cache compression methods, KVQuant, H$_2$O, and RocketKV, together with an uncompressed FP16 configuration, on four datasets covering mathematics, science, commonsense reasoning, and reading comprehension. Across a wide range of latency and peak memory constraints, MetaKV consistently outperforms the best static configuration, improving constrained success rate (CSR), the fraction of prompts answered correctly while satisfying both constraints, by approximately 0.07 on average and up to 0.135. These results demonstrate the benefit of adapting KV cache compression to individual prompts and latency-memory constraints. Code is available at https://github.com/MichaelWang0505/MetaKV.git

Caching systems often rely on simple eviction policies such as Least Recently Used (LRU) and Least Frequently Used (LFU), which perform well in complementary request regimes. Recent policies such as LeCar and Cacheus combine LRU and LFU using ideas from the experts problem in online learning. Specifically, upon a miss, they randomize between the two eviction rules using probabilities derived from scores updated by tracking the history of past evictions. While these policies exhibit strong empirical performance, it remains unclear whether they are guaranteed, on every request sequence, to perform asymptotically as well as the better of LRU and LFU, i.e., whether they achieve sublinear regret with respect to this benchmark. We first show that LeCar suffers linear regret against an oblivious adversary, even with unbounded history. We then propose H-MC, a Hedge-based mixture of virtual LRU and LFU caches that preserves Hedge's selection probabilities, and hence its regret guarantees, while minimizing the switching cost among all joint selection rules with these marginals.

A long-lived KV cache must be compressed before the queries that will read it exist. Selection by observed attention collapses there: on a NoPE-MLA model, H2O and SnapKV retrieve 0.00 and 0.33 of needles at 8x compression, because a token's importance has not yet been observed. VestigeKV instead derives a sparse attention pattern from a signal the cache already carries, occupying the sparse-attention literature's one unoccupied quadrant: training-free and query-independent. In NoPE-MLA the 64-dimensional decoupled branch is a vestige of RoPE that training repurposes into a salience channel; reading 11% of each row, it partitions the cache into an attended tier and a GPU-resident archive that no row ever leaves, reachable each step by a certified, query-adaptive trigger. Nothing is trained and cache rows are never quantized, so every quality effect attributes to selection and scheduling. On Kimi Linear 48B, retrieval holds at 1.00 under 8x and 0.96 under 32x from 8k to 65k context, with zero gap to full-row selection, and the recall tier holds 128x at 1.00 (8k). Both tiers stay on the GPU, so the win is speed, not memory: the per-step scan reads ~26% of the bytes dense attention would, and on a two-node sglang deployment the crossover sits at ~40k context, reaching 1.18x at 256k and 1.39x at 496k. The mechanism is exclusive to NoPE: the identical operator on a RoPE MLA collapses to 0.08, query-independent salience exists only without rotation, and query-universal exact merging is provably impossible under RoPE. All thresholds were frozen before their data; 20 archived verdicts and 8 closed routes accompany the paper.

Decoding-time KV cache compression research focuses heavily on designing better token scoring functions, while the temporal rule that aggregates scores across decode steps is often treated as an implementation detail. Under aggressive KV compression, we find that exponential-moving-average (EMA) aggregation makes approximately order-preserving scorer modifications largely indistinguishable at the eviction-set level. Value-norm and entropy variants remain highly correlated with attention and produce nearly unchanged retention sets, whereas KeyDiff, key norm, recency, and a learned scorer alter the ranking and degrade substantially. We associate this stability with the evaluated aggregation, which couples layer weighting and temporal retention. Building on this observation, we introduce InertiaKV, an EMA-based decoding-time eviction method, and InertiaKV-Lazy, its periodic-refresh variant, which yields 1.34-1.46x decode throughput relative to full refresh InertiaKV. We also study Score-Free decoding as a separate empirical operating point: it scores the full context once at the first decode step, freezes that ranking, and incurs an average quality change of +0.03 while removing all subsequent scoring. Across six open-weight backbones and the LongBench, LongBench-v2, and RULER benchmarks, the results identify temporal aggregation and ranking preservation as distinct, consequential design factors; they do not imply that scoring quality is irrelevant in general.

Large language models achieve superior performance on tasks that require extended reasoning, but long chains of thought make the KV cache a severe memory bottleneck. Existing KV cache compression methods share one paradigm: score each cached token by some estimate of how much it will matter later, and keep the top-scoring ones. We show that the selection signal contributes almost nothing. Random Attention keeps the prompt and evicts uniformly at random within each attention head, computing no score at all; across four models and six reasoning tasks it matches the strongest prior evictor while serving 32-43% higher throughput than it in vLLM deployment. Controlled experiments explain this by showing that 1) the prompt is the fragile part of the cache, and most of the gap between selectors is just whether their selection signal happened to keep it; 2) the reasoning trace protects itself against eviction with redundancy at two levels, in the text (the model restates what it still needs as it works) and across attention heads (each keeps its own copy of the trace), so once the prompt is safe, a random draw retains enough copies of what the model still needs, and no score is required to pick them. Our code is publicly available at https://github.com/SalesforceAIResearch/Random-Attention.

Large language models (LLMs) face severe memory bottlenecks in long-context inference due to the linearly growing size of key-value (KV) caches. Existing KV cache compression techniques typically rely on simple heuristics, overlooking the distinct functional roles of different attention heads. We present SGD-KV (Summarization-Guided KV Cache Compression), a head-aware framework that leverages a novel chunk-summarization diagnostic task to systematically identify and prioritize attention heads specialized in hierarchical information aggregation. Experiments on Qwen2.5-7B-1M and Qwen3-32B across diverse long-context benchmarks demonstrate that SGD-KV achieves state-of-the-art performance with contexts up to 1M tokens, while reducing KV cache memory usage by up to 75%. Our findings show that strategically allocating the KV cache budget based on the summarization score distribution of attention heads yields a superior efficiency-accuracy trade-off for long-context inference.

Long-context inference retains a growing key--value (KV) cache during decoding, which consumes substantial GPU memory and can reduce generation throughput. This bottleneck remains in hybrid language models because their residual global-attention layers can dominate context-dependent cache demand. We study how to allocate this state under an aggregate KV-residency budget. We introduce HeadWiseKV, a training-free framework that compresses the residual global KV caches of hybrid language models while preserving their native local, recurrent, and linear paths. It assigns each physical KV head a static, multilevel history window, making cache demand predictable before serving. We formulate this allocation as a restricted operational rate--distortion problem and propose SeqCalib as the core policy-generation algorithm in HeadWiseKV. SeqCalib processes layers in execution order and conditions each decision on the lower-layer policy used at deployment, thereby accounting for interactions across depth. A grouped-cache runtime materializes the selected policy as actual per-head KV residency rather than a mask over a full cache. We evaluate downstream quality across four hybrid long-context models and study physical residency and serving behavior on Qwen3.6-27B. HeadWiseKV retains near-Full-KV RULER and LoCoMo quality across the evaluated models. In the fixed-model systems study, it reduces sampled peak device memory by 8.59% at a 112K context length and extends the largest verified successful context from 114K to 161K.

Spoken dialogue state tracking recovers slot-value pairs from speech, where ASR errors concentrate in entity values and persist across turns, making it both a generation and an editing problem. A strong per-turn text editor corrects much of this but, operating on the transcript alone, leaves three recoverable errors: a value predicted inconsistently across turns, an omitted slot, and a value the audio does not support. We present AVERT, which scores each candidate value by combining cross-turn agreement with a trained audio-conditioned verifier and resolves the three error types with three operators, vote, add, and swap, each restricted to the slots where its error is common. On SpokenWOZ, a base speech-LLM reaches 33.04 JGA, a text editor 38.34, and AVERT 40.13, without retraining either. This is in the range of a 1B end-to-end system that consumes the full spoken history (39.32), though AVERT uses two 1B decoders rather than one. The audio verifier contributes a statistically significant gain, and restricting each operator to a selected slot subset matters: removing it lets unrestricted voting overwrite correct categorical values and fall below the editor.

Long-horizon large language model (LLM) agents require context assembly: the runtime must decide what to include in each prompt, in what order, and when to compact history under a hard context-window budget and a byte-sensitive prompt cache. In production agentic systems, this logic is scattered across prompt builders, ad hoc compaction routines, cache-break workarounds, and per-provider shims. We argue that context assembly is structurally isomorphic to query execution in a relational database: both execute under a hard budget, exploit a tiered cache, and leverage statistics. We adopt this discipline in ContextPipe: a five-phase pipeline (Plan Bind Optimize Execute Feedback) backed by a structured data-source catalog, a deterministic cache-aware optimizer, and an EXPLAIN ANALYZE trace. We show that context in ContextPipe is auditable, replayable, and failure-isolated. A preliminary evaluation using the SWE-bench Pro Qutebrowser subset shows that, compared with the append-only context construction policy, ContextPipe reduces total token volume by 31%, LLM calls by 23%, and response time by 9%, at the cost of a lower KV cache-hit ratio.

LLM agents that cache recovery suggestions from API errors can skip re-derivation in later episodes, spending fewer tokens and fewer model calls on constraints they have already learned. Server-side data drift turns those cached fixes into silent failures, and the usual remedy, re-deriving on every episode, gives the savings back. We introduce invalidation contracts, a protocol layer that attaches version stamps and cacheability hints to every recovery suggestion so the client can evict stale entries without trial and error, and keep the rest. The contract decomposes realized savings into two independent factors: validity, the fraction of cached suggestions that remain correct after a drift event, and compliance, the fraction the planner applies on the first attempt. Validity depends only on the protocol and is vendor-independent. Compliance depends on the planner model: identical wire bytes yield 100% first-try compliance on Claude Haiku 4.5 and 11% or below on Claude Sonnet 5, which exhibits input-schema conservatism, refusing fixes that add fields the original request did not contain. We evaluate across seven models, three serving paths, two domains, and approximately 9,400 episodes. Row-level invalidation raises compliance by 0 to 66.7 percentage points across the seven models, 55.6 to 66.7 on three, and recovers 29-33% of baseline token cost on four of seven models, while table-level invalidation destroys co-located entries and drops post-drift first-try rates to 0% on five of seven. Eviction precision is 1.00 at row granularity on every model under the row-level oracle of Section 4.1. The contract adds 15% to response payload. Version-stamp validity is deterministic by construction and produced identical results across every model and serving path, with zero contract failures in the entire evaluation.

Retrieval-augmented generation systems can precompute and store key-value caches of retrieved documents to avoid re-encoding context at every query. Quantizing these caches further reduces storage, but no prior work asks whether compression damages faithfulness, whether responses remain grounded in the retrieved evidence. Faithfulness and accuracy are not equivalent: a model can produce a correct answer that is no longer supported by the context it was given. We evaluate Qwen2.5-7B-Instruct under INT8 and INT4 quantization on RGB and HotpotQA, measuring both accuracy and faithfulness with a hallucination detector, NLI entailment, and an LLM judge. INT8 is near-lossless across both metrics. INT4 reduces accuracy and, more critically, even among answers that remain factually correct, over 90% of faithfulness changes are negative, i.e., accuracy metrics are blind to this regression. The harm grows under noisy retrieval and with more retrieved chunks. Faithfulness must be audited before compressed caches are deployed.

Modern large language model (LLM) serving systems increasingly operate over repeated or shared context, yet each model typically performs its own prefill computation even when another model has already processed the same input. Existing KV-cache reuse mechanisms substantially reduce redundant computation within a single model, but generally assume that the producer and consumer of a cache are identical. We study \emph{cross-model KV sharing}, which translates the KV state produced by a source model into a representation that can be consumed by a different target model, including models that differ in scale, architecture, attention configuration, tokenizer, and model family. We evaluate the approach in both within-family and cross-family settings. For Qwen2.5-7B $\rightarrow$ Qwen2.5-1.5B, translated KV states improve LongBench2 accuracy from 27.59% to 34.48%, a gain of 6.89 percentage points over the native 1.5B baseline, while reducing handoff cost relative to native target prefill. For the cross-family Qwen2.5-1.5B $\rightarrow$ Gemma-2-2B setting, KV handoff reduces target-side prefill cost by up to 67.05% at 4K context length while maintaining decoding perplexity close to native-model baselines. In a more heterogeneous Llama3.1-70B $\rightarrow$ Qwen2.5-7B setting, cross-family handoff achieves 44.0% accuracy compared with 45.7% for native Qwen2.5-7B inference, while reducing measured latency from 899ms to 138ms. These results provide initial evidence that KV states can serve as transferable computational representations rather than strictly model-local caches, and motivate \emph{context mobility} as a systems abstraction for reducing redundant prefill across heterogeneous LLM and multi-agent inference workflows.

Language models can answer from precomputed memory, a model's saved reading of a body of material, reused across requests instead of read again at each. This paper maps where that practice preserves correctness and the conditions under which it fails. Across experiments on Llama-3.1-8B-Instruct using both saved key-value caches and trained compressions of them, precomputed memory degrades when assembled from separately prepared parts, stays current only through rebuilds costing a large fraction of full preparation in our measurements, and ignores corrections served beside it conditional on phrasing. If precomputed memories can be served alongside one another, be cost-efficiently rebuilt, and be superseded by new information arriving in real-time, they can serve as a way to avoid re-feeding context to a model over repeated queries. The implication of our results for a deployed system that deals with a variety of queries is that precomputed memories are best rebuilt on the cadence at which new information changes what the memory was originally computed from. Both warm-rebuilding trained compressions of key-value caches and serving specifically-phrased updates beside a memory, as pasted text or injected cache state, show particular promise for keeping precomputed memories current, the latter as an interim measure between rebuilds, and we measure the cost and name the remaining questions associated with each.

Inference with transformer-based large language models (LLMs) is often limited by the memory-bound KV cache and quadratic attention cost. State-space models (SSMs) mitigate this through linear attention and fixed-size recurrent states, but their large dense linear projections remain computationally expensive even after quantization. We introduce a method that induces sparse neural activity in heavily quantized linear-attention models with minimal performance loss. Activations below a per-projection trainable threshold ($\pm Δ$) are nullified while preserving crucial outliers, achieving comparable performance to dense models with up to 4$\times$ fewer effective arithmetic operations. Targeting a multi-core, multi-chip neuromorphic platform, where event-driven execution converts unstructured sparsity into throughput at both the compute and communication levels, a capability GPU architectures fundamentally lack, we project up to 37$\times$ higher throughput and 16$\times$ lower power versus edge GPU inference of a comparable transformer-based model, and up to 5.4$\times$ improvements over the non-sparsified baseline. These results position sparse, quantized linear-attention models as a natural fit for deploying LLMs on event-driven multi-core platforms.

Hybrid large language models interleave full-attention layers with linear-attention layers to reduce the cost of long-context inference. This structure complicates prefix caching: full-attention key-value caches are token-addressable, whereas linear-attention layers maintain recurrent states that cannot be rolled back to arbitrary prefix boundaries. Existing hybrid prefix caching methods address this mismatch by storing recurrent-state checkpoints. As a result, token-level matches are directly usable only at positions aligned with stored checkpoints, constraining prefix reuse to a discrete set of boundaries. We present Tail-Replay, a prefix caching mechanism that enables unconstrained token-level prefix reuse in hybrid large language models. The key insight is that linear-attention mechanisms such as Gated DeltaNet can be viewed as a structured, lossy compression of the input prefix: gated recurrent updates progressively attenuate the contributions of earlier inputs. Consequently, the recurrent state of a matched prefix can be well approximated by replaying only a short, recent suffix of that prefix. Tail-Replay exploits this property by caching the exact full-attention key-value cache while omitting recurrent-state checkpoints. On a cache hit, it reconstructs the linear-attention states by replaying a short, recent suffix of the matched prefix. As a result, the reuse boundary is determined by the shared tokens rather than by recurrent-state checkpoints. We evaluate Tail-Replay on three Gated DeltaNet-based hybrid models using the LongBench and RULER benchmarks. With only a 5--10% replay budget, it retains 92.8--99.9% of full-prefill quality on LongBench and RULER. For serving efficiency, we evaluate time-to-first-token speedups across multiple matched-prefix lengths---8K, 16K, and 32K. The speedup grows with prefix length, reaching $9.1$--$14.3\times$ over full prefill at 32K.

Single-GPU deployment of 70B-parameter language models on an NVIDIA GPU is constrained by device memory, long-context throughput, and engineering integration cost. We cast single-GPU inference as a budget-aware design problem over these three axes and study how pruning, quantization, and KV-cache compression interact under realistic execution. Controlled ablations show that layer-wise pruning makes weight quantization more robust. KV-cache sparsification complements INT8 KV quantization by reducing memory without hurting decoding speed, while static vector quantizers often conflict with dynamic caching. Guided by these coupling results and explicit budget tracking, we assembled a practical pipeline and compressed a 70B model to about 33 GB, sustained about 57 tokens/s on 10k token prompts on a single A40, and kept absolute accuracy within 5% on common and reasoning benchmarks. We contribute design rules and a reproducible evaluation protocol that jointly report quality, memory, and end-to-end speed, and we provide a foundation for automated pipeline search under realistic single-GPU constraints.

Real-world image generation often involves multi-turn editing, where users iteratively modify small regions while most image content remains unchanged. However, existing diffusion transformer (DiT)-based editing pipelines recompute the entire image at every turn, causing substantial redundant computation. Existing DiT acceleration methods further ignore semantic correspondence across prompts, leading to unnecessary recomputation or unsafe reuse that harms editing quality. To address this, we propose RegionCache, a semantic-aware reuse framework for multi-turn image editing that selectively reuses diffusion states from unchanged regions. RegionCache detects reusable regions through semantic overlap between consecutive prompts and cross-attention localization, and adopts an adaptive reuse schedule based on prompt similarity and contextual consistency. Experiments on PixArt-alpha demonstrate that RegionCache achieves 1.43x--2.55x end-to-end speedup while maintaining comparable image quality. Code is available at https://github.com/hebutBryant/RegionCache.

Unified multimodal models jointly support understanding and generation, but incur substantial redundant computation across tokens, layers, and generation timesteps. Through token-importance probing, we identify an asymmetric core-expansion structure: understanding exhibits a stable importance component, while generation largely shares this component but requires progress-dependent corrections. We therefore propose CE-Router, which uses a task-shared core scorer and progress-conditioned generation expansions, optimized through generation decomposition and cross-task core alignment. At inference, CE-Router compacts token computation and supplies a learned routing signal to Unified Computation Scheduling, which coordinates layer skipping, FFN pruning, diffusion-head cache reuse, and denoising-step early exit. Experiments on two representative UMM architectures demonstrate consistent quality--efficiency improvements across both tasks, retaining 98.03% of dense understanding performance with a 1.93$\times$ end-to-end inference speedup.

Long-context inference is bottlenecked by the memory footprint of the key-value (KV) cache, especially for small models under tight resource budgets. Existing KV cache eviction methods score tokens using the model's attention distribution or, in attention-free variants, each key's distance from a global reference point. Using a controlled leave-one-out probe, we find that attention magnitude is unrelated to a token's causal contribution to the answer (Spearman $ρ=-0.004$), challenging the premise behind dominant eviction methods. We introduce TwinKV, a training-free, attention-free redundancy signal that detects whether a token's key has a near-duplicate elsewhere in context. Rather than replacing existing policies, TwinKV acts as a composable repair pass: given a policy's fixed retained set, it identifies evicted tokens with no surviving duplicate (\emph{orphans}) and retained tokens whose information is duplicated elsewhere (\emph{redundant donors}), then swaps them while preserving the original budget and scoring rule. We compose TwinKV with four recent eviction policies across LongBench, LooGLE, RULER, and a short-context MMLU-Pro no-harm control at compression ratios ${0.3,0.5,0.7}$. On Qwen3-4B, TwinKV improves a majority of configurations for two policies, is near-even for a third, and helps only a minority for a fourth adaptive baseline already near a performance ceiling; gains across the three non-ceiling policies are smallest at the loosest ratio. On RULER with Llama-3.2-1B, however, that fourth policy improves in every evaluated cell because its Alone score leaves substantial room to improve. More broadly, Llama-3.2-1B shows a smaller average LongBench gain but a higher fraction of improved cells on LongBench and LooGLE than Qwen3-4B, plus a clean RULER win. We also identify few-shot classification exemplars as a task structure where TwinKV does not help on either model.

Key-value (KV) caching is essential for efficient autoregressive large language model (LLM) inference, but the cache grows linearly with context length, increasing storage and decoding costs. KV cache compression mitigates this cost by retaining only a subset of cached tokens. This challenge is particularly important for multi-step LLM agents, where a query expands into trajectories of reasoning, tool interactions, and retrieved observations. Existing pruning methods typically treat the cache as a flat token stream and rank tokens by recency or attention saliency. This creates a mismatch between the unit of compression and the unit of reasoning: token-level pruning removes individual entries, whereas useful information in multi-step agents is often organized into reasoning steps with uneven and delayed importance. Consequently, an early observation or intermediate decision may receive little recent attention yet remain essential for later evidence synthesis. We term this failure mode Reasoning Continuity Disruption.These observations motivate KV cache compression that jointly considers token- and reasoning-step-level information. StepKV addresses this goal by treating reasoning steps as first-class retention units. It associates cache entries with their generating steps, estimates step utility from trajectory-derived signals, and combines this utility with token-level saliency. The resulting scores globally rank prunable tokens, from which StepKV retains the top-scoring entries under a target budget. StepKV thus provides a step-centric perspective for agent KV cache compression. Across multi-hop QA and long-horizon web reasoning tasks, StepKV sustains accuracy under low KV budgets where token-level baselines degrade sharply, offering a more robust efficiency-accuracy trade-off for multi-step agent inference.

KV-cache eviction can do more than compress. In long-context LLMs, keeping only some cached tokens sometimes matches or exceeds full-cache accuracy, because many redundant prefill tokens otherwise dilute attention away from the tokens that carry the answer. This benefit is not uniform, and evicting the wrong tokens can drop accuracy to zero on tasks that require precise retrieval, so the useful question is not only which tokens to keep but also whether to evict this input at all. We show that one label-free number computed from the prefill attention, the drop between early and late layers in how much attention heads agree on which tokens to read, predicts per input, before any decoding, which of the two cases an input falls under. We build this into PAGE (Partition-Aware Gated Eviction), a wrapper that runs any SnapKV-style evictor when the drop is large and keeps the full cache when it is small, with no training, labels, or fine-tuning. PAGE is a safety mechanism rather than a compressor, so we measure it by the failures it prevents. It cuts the harm rate on capacity-bound inputs from 0.75 to 0.026, and on multi-key retrieval with Mistral-7B plain SnapKV falls from 99% to 0% as the budget shrinks, while PAGE holds it at 89%. Elsewhere, it passes the base evictor through unchanged, which is the intended behaviour and is what we observe in 8 of 16 cells. Code is available at https://anonymous.4open.science/r/PAGE-018239.

Large language model serving faces a critical memory bottleneck: the KV cache grows with sequence length and batch size. PagedAttention uses fixed-size memory blocks to reduce allocator-level fragmentation, but recent KV eviction algorithms operate at a token granularity finer than block-level management. This mismatch causes intra-block fragmentation, leaving a large fraction of allocated KV memory unreclaimable. We present vToken, a lightweight token-level virtualization layer that decouples logical token liveness from physical block placement. vToken maintains a stable logical token view through token-table indirection and realizes physical reclamation by repacking live tokens asynchronously. The design preserves PagedAttention kernels and CUDA Graph compatibility. We implement vToken in vLLM and evaluate it with H2O, Random, and Scissorhands across models. Compared with a paired Naive-Evict baseline, vToken reduces retained KV blocks per request by 27.2%--72.3% and improves SLA-constrained throughput by up to 1.37$\times$. Under a constrained active-KV budget, it extends the maximum feasible concurrency by up to 2$\times$, while reducing the per-policy integration footprint from 500+ lines to under 50.

As edge-side vision services continue to expand toward low-latency, high-throughput scenarios, reducing the inference cost of vision models without sacrificing reliability has become a central concern. Existing semantic caching methods largely rely on empirical similarity thresholds; while such thresholds improve hit rates, they tend to introduce silent misclassifications near decision boundaries. To address this, we propose \texttt{LipCache}, a certified semantic caching framework for image classification. Without modifying the existing deployed main model, \texttt{MainNet}, the framework introduces a lightweight network, \texttt{GuardNet}, that maps inputs into a low-dimensional feature space subject to a Lipschitz constraint. It then computes a per-sample certified reuse radius from the local classification margin and the spectral norm of the classification head. At runtime, a cached result is reused only when the query feature falls inside the certified reuse ball; otherwise, the query falls back to \texttt{MainNet}. Thus, cache hits are transformed from empirical threshold tests into geometric certification decisions with explicit theoretical boundaries. Across standard image classification tasks like CIFAR, Tiny-ImageNet, and SVHN, \texttt{LipCache} achieves a measured speedup of up to $1.65\times$ with limited end-to-end accuracy degradation, while all accepted cache hits satisfy the \texttt{GuardNet}-side certified-consistency condition. Furthermore, an enhanced \texttt{GuardNet} training recipe substantially improves cache hit rates in the Tiny-ImageNet multi-class extension while maintaining a certified-consistency rate of $100\%$. These results demonstrate that per-sample certified reuse can reduce main-model fallback while preserving theoretical consistency, providing a feasible approach to reliable cache-assisted inference at the edge.

Diffusion models have achieved dominant performance in visual generation but suffer from substantial inference overhead. While cache-based acceleration has emerged as a promising solution, existing policies rely on local similarity heuristics, which we identify as being significantly misaligned with final generation quality. This discrepancy stems from the non-uniform propagation and accumulation of errors along the denoising trajectory. To address this, we propose Global-Impact Cache (GCache). We first establish a rigorous theoretical characterization of the error propagation upper bound. Recognizing that this bound can be overly conservative for complex, highly non-convex diffusion models, we further reparameterize the propagation exponent with a Bernstein form and reformulate cache policy search as a bilevel optimization problem. In detail, GCache identifies an optimal reuse policy in the inner objective while aligning the error-weighting function with generation quality loss in the outer objective. This framework effectively reconciles theoretical rigor with empirical performance, learning to prioritize computation where it most impacts visual fidelity. Extensive experiments demonstrate that GCache consistently outperforms prior caching strategies on both video and image generation. Notably, on the state-of-the-art Wan2.1 video diffusion model, GCache maintains a 2.17x speedup while significantly enhancing generation quality, reducing LPIPS from 0.1095 to 0.0316.

Retrieval-Augmented Generation (RAG) repeatedly prefills identical text chunks across queries, incurring redundant computations. Position-Independent Caching (PIC) mitigates it by reusing precomputed Key-Value (KV) across positions, but its efficiency is constrained by the large volume of text tokens. Rendering text chunks as images can compress the text into fewer visual tokens, but the rendered-image PIC suffers more severe quality degradation than the text PIC. This representation-specific gap primarily arises from contextual mismatches across independently compiled caches and the loss of fine-grained textual evidence during visual compression. Existing PIC repair methods mainly address the former through selective recomputation, but they incur online computation and cannot recover lost textual details. We propose QV-PIC, a query-aware dual-resolution PIC reuse framework guided by model-native templates. Offline, QV-PIC compiles visual caches under the model's native chat-template prefix, improving PIC quality without online recomputation. Online, it preserves global context with low resolution and restores fine-grained textual evidence within a high-resolution budget by cumulative query relevance scores, retaining the efficiency benefit of visual compression. Across six tasks, QV-PIC improves average F1 by 21.6 points over vanilla rendered-image PIC, closes the gap to vanilla text PIC, and surpasses optimized text PIC by 2.58 F1 while reducing TTFT by 17.2%. Relative to full prefill, it cuts TTFT by 83.8%.

World action models (WAMs) condition robot actions on predicted futures, but iterative video rollout increases deployment latency. We ask whether action generation requires the evolving rollout trajectory or only its future representation. Across four WAMs on all 40 LIBERO tasks, paired closed-loop interventions show that masking or reassigning future-cache values changes execution and reduces success, indicating sensitivity to future values and their assigned positions. For Joint and Cosmos-2, however, replaying one fixed final-clean key/value (K/V) cache nearly preserves unmodified execution, with $1.7$ to $1.9$cm end-effector average displacement error and $97.9\%$ to $98.2\%$ success. This separates cache consumption from production: these models can reuse a fixed cache but still require iterative rollout to construct it. We therefore propose RIFT (\emph{Rollout-free Imagination via Future Tokens}), which uses learned anticipation tokens to construct a complete future K/V cache in one backbone pass while retaining the original future-read interface. On LIBERO, RIFT achieves $98.8\%$ success, close to rollout-based Joint, IDM, and LingBot-VA at $98.4\%$ to $98.6\%$, while reducing action-chunk latency by $68.2\%$ to $89.1\%$. On RoboTwin2.0, RIFT reaches $92.9/92.6\%$ on clean/randomized scenes, the highest observed among the evaluated methods. These results support rollout-free future conditioning without iterative video generation at deployment.

Vision-Language-Action(VLA) models map camera images and language instructions directly to motor commands through a single autoregressive transformer. In real-time control, they still spend substantial compute recomputing key-value(KV) representations for visual tokens that barely change across neighboring frames. Recent work such as VLA-Cache reduces that cost by reusing KV states for visually static patches, but its policy relies only on observation-space heuristics and does not account for the model's own uncertainty. We propose Gated VLA-Cache, a lightweight, training-free extension that augments visual-similarity caching with neural introspection. The method monitors the logit margin between the top two predicted action tokens, a zero-cost confidence signal available during decoding. When the margin drops below a threshold, the cache is invalidated and a full recompute is triggered. Evaluated on four LIBERO benchmark suites with both OpenVLA and OpenVLA-OFT, Gated VLA-Cache improves reliability when blind caching hurts. On LIBERO-Goal and LIBERO-Long, it recovers over 100% of the lost accuracy while retaining 80% of the compute savings.

Edge LLMs must preserve inference continuity when a user hands over between edge nodes, requiring key-value (KV) cache transfer to the target node. However, simultaneous handovers saturate the backhaul, preventing full cache delivery within the mobility-imposed transfer window. Rather than allocating bandwidth as if all cache entries were equally valuable, we order each user's KV cache by importance and transmit only its most informative fraction, turning token-level sparsity into communication savings. We cast the transfer as a multi-user backhaul allocation problem that maximizes average accuracy across users. Each user's partial-cache accuracy serves as its utility: a sigmoid that fits measurements on the RULER benchmark with $R^2>0.99$ across models and context lengths. Because importance ordering front-loads the high-value entries, the concave region of the accuracy curve spans nearly the entire cache. Our proposed allocator keeps served users within this region, making each per-slot allocation problem convex. The optimum is derived via a closed-form weighted water-filling solution that generalizes information-theoretic water-filling and enables online scheduling. The proposed allocator attains over 93.7% average accuracy in a 500ms transfer window, within 0.5pp of the full-cache ceiling, and reaches 98.2-99.5% of a clairvoyant upper bound.