生成模型与LLM推理优化

In agentic LLM services, a session calls an external tool, waits for it, and re-arrives to continue inference. Statically compiled NPU serving can fix the batch bucket set, the maximum batch size, and the number of KV cache slots at compile time. We define such an environment as a compile-time-static serving substrate and analyze the execution-time cost that tool waiting and re-arrival incur in it. On a single LLM instance, we run synthetic workloads following a measured tool waiting time distribution and compare, on the same inputs, a baseline configuration with settings {1, 2, 4, 8}, 8, and 8 against configurations that change some of them. Because re-arrival times differ across configurations, we build a simulator that replays request processing in time order, select the candidate with the lowest predicted cost among 2,077 configurations, and validate it on new inputs. We identify three mechanisms: discrete batch alignment, KV cache survival, and prefill interference. At a concurrency of 6, absent from the bucket set, tool waiting lowered the padding ratio (0.235 to 0.120) yet increased decode execution time 1.51-fold, so padding alone did not indicate cost. On new inputs at a concurrency of 8, where the baseline reused KV in 9 of 24 re-arrivals, enlarging the maximum batch size alone cut execution cost by 8.25%, and the selected configuration, which also adjusted the bucket set, by 9.72%. Where 17 of 18 re-arrivals were already reused, the effect was 0.59%. Compile-time configurations should thus be selected by diagnosing KV reuse loss and the resulting change in execution.

Serving long documents to a Large Language Model (LLM) repeatedly is expensive: computations grow with context length, and the memory footprint of the key-value (KV) cache balloons. Compressed KV (CKV) representations aim to mimic the cache of a document and are typically computed once and for all, ahead of inference time. Methods to obtain CKVs range from drop mechanisms that reduce their number of columns, to learned approaches. Among the latter, Cartridges have emerged as a leading compression method, learning compact KV representations through distillation on relevant Q/A pairs. While existing evaluations focus primarily on whether Cartridges and other CKVs yield approximately similar responses to document-related, on-context queries, we investigate the crucial deployment question of whether they can handle off-context queries, something the native KV representation is particularly good at, thanks to the mechanics of attention. We observe a fundamental trade-off: while Cartridges perform better for on-context queries, heuristic-variants preserve better the original LLM's ability to operate off-context. We measure this through their capability to avoid context contamination in their response, retain general knowledge, and follow instructions. We propose Cartridges++, simple modifications to cartridges that retain off-context abilities at small or negligible cost. The router variant decides at inference time whether the query should use the learned long-context memory, while the data-mixing variant allocates a small fraction of training Q/As to queries outside the reference long document. Our study shows that assessing CKVs on document utility alone can mask substantial degradation in broader model capabilities, yet those issues can be fixed with benign changes to CKV inference or training.

Spexis is a multi-GPU LLM inference framework that improves the efficiency of pipeline and tensor parallelism through speculative parallelism. Rather than using speculative decoding only to accelerate token generation, Spexis runs speculation in parallel with normal execution, introducing a new parallelism axis without increasing KV-cache memory usage. This improves memory efficiency and helps mitigate the bottlenecks of multi-GPU inference. Spexis further uses lookahead scheduling to predict speculation quality and future memory pressure, allowing it to reduce wasted speculation, KV-cache eviction, and recomputation. Built on top of vLLM, Spexis largely improves serving performance across a range of GPU configurations, achieving speedups of up to 34% over a baseline that uses the optimal combination of pipeline and tensor parallelism. Spexis's source code is publicly available at https://github.com/mlsys-seo/spexis.

Multi-document retrieval-augmented generation (RAG) requires a language model to process multiple retrieved text chunks before answering a question. Precomputing each chunk's KV cache independently and concatenating the caches when the chunks are retrieved can accelerate this step. However, the assembled cache lacks cross-chunk attention information, reducing answer quality. Selective recomputation methods recover the missing cross-chunk context by rerunning the target LLM on selected tokens, incurring substantial online computation. We introduce CacheRepair, a lightweight network that learns the difference between independently computed KV caches and those produced by processing the chunks together. The network combines compressed KV features with token embeddings and uses attention that is bidirectional within each chunk and flows from earlier to later chunks. Each repair block receives the compressed cache features, and the predicted residual is added to every document token's cache. Each repair network is trained for a specific frozen target LLM on a generic retrieval corpus and reused across downstream datasets. Our analysis shows that repair reduces KV errors both near chunk boundaries and throughout chunk interiors. Evaluation across three target LLMs and four downstream datasets places CacheRepair on the measured answer-quality-latency Pareto frontier in eleven of twelve model-dataset combinations. Reported time to first token (TTFT) includes online cache transfer and repair. Across all twelve combinations, the largest repairers achieve 1.69-4.61$\times$ speedups in median TTFT over full prefill and improve mean F1 by 2.1-26.1 percentage points over direct cache reuse.

On-device large language model (LLM) serving is a cornerstone of local-first personal intelligence, offering users data sovereignty, strong privacy guarantees, and freedom from cloud API latency and cost. Although KV caching is widely used to reduce latency in long-context inference, existing designs were primarily optimized for cloud GPUs with dynamic execution environments and abundant memory bandwidth. These architectural assumptions do not hold on mobile NPUs, where computation graphs must be statically compiled and both memory capacity and I/O bandwidth are severely constrained. In this work, we present a compute-storage co-design for mobile-centric prefix and non-prefix KV reuse. We first propose an intra-graph mechanism that maps selective KV recomputation onto static NPU graphs, reconciling algorithmic dynamicity with NPU staticity. We further develop an inter-graph scheduler to optimize chunk merging and minimize padding with dynamic programming. To address mobile bandwidth limitations, we introduce a hierarchical KV manager featuring a tree-hash-semantic hybrid structure, along with cost-aware prefetching and eviction policies. We also build a two-dimensional pipeline that overlaps KV loading, rerotation, and storage with NPU execution, hiding data-movement latency. Experiments across representative on-device workloads and LLMs show that our design reduces time-to-first-token (TTFT) by $40-60%$ compared with no reuse and prefix-only caching.

Reusable prefix key-value (KV) caches can outgrow GPU memory in large language model (LLM) serving. A memory-semantic flash hierarchy offers SSD-backed capacity with a limited fast tier, but a logical KV hit is not necessarily ready for GPU retrieval. Demand staging exposes SSD latency, whereas immediate staging can reserve fast-tier capacity long before retrieval begins. We present TempoKV, a timing-aware resource-commitment layer that separates early knowledge of reuse from the acquisition of staging resources. It records reusable-KV hits as metadata-only claims and requests commitment when the runtime-estimated time until retrieval falls to the storage-estimated time needed to make KV resident and protected against eviction. These estimates adapt to runtime progress and staging state, while commitment remains subject to available protected capacity. We implement TempoKV in vLLM and LMCache on an SSD-backed CXL memory device without changing request scheduling. Across two models and three prefix cache ratios, TempoKV reduces protected fast-tier byte-time per request by 63-91% versus immediate staging while retaining much of the serving benefit of advance staging. In a fast-tier capacity sweep, output throughput and p95 time to first token (TTFT) remain nearly unchanged as capacity decreases from 100 to 25 GiB. Compared with unmodified LMCache's Device-DAX L1 configuration, TempoKV reduces p95 TTFT by up to 48.0% and increases output throughput by up to 27.8%.

Mixture-of-experts (MoE) models activate few experts per token, yet batched decoding can access nearly the entire expert pool, making expert-weight traffic a major bottleneck. Expert pruning reduces this traffic, but conventional approaches also prune compute-bound prefill, sacrificing model quality for little throughput benefit. We present SlimWise, a serving framework that tailors the expert pool to each inference phase. SlimWise performs prefill with the full model and decode with a pruned model that directly reuses the prefill-generated KV cache without conversion. Across two MoE backbones and three pruning criteria, this training-free KV cache handoff substantially narrows accuracy gaps relative to the full model in many settings. We also show that benchmark accuracy can conceal substantial pruning-induced changes in generation length. To address these distortions and residual accuracy loss, SlimWise introduces a low-cost distillation stage that trains the decoder to continue from full-model KV caches while updating only a small subset of parameters. Implemented in vLLM, SlimWise supports both prefill-decode (PD) disaggregation and PD-colocated serving. On Qwen3.6-35B-A3B, SlimWise improves decode throughput by up to 1.81x at 50% expert pruning with minimal accuracy loss.

Long-context LLM serving is increasingly bottlenecked by decode, where large KV caches limit batch size and underutilize GPUs. Sparse KV cache offloading expands effective capacity by storing most historical KV blocks in CPU DRAM and recalling only selected blocks on demand. However, we find that existing offloading systems shift the bottleneck to CPU-GPU recall I/O: recall volume varies widely across layers, decode steps and requests, while headwise sparse selection fragments recalls into many small PCIe transfers. This paper presents PulseInfer, an I/O-centric sparse KV cache offloading system. PulseInfer hides variable recall latency with interruptible layer-wise scheduling, adapts offloading decisions with IO-Adaptive Offloading Admission, and coalesces fragmented transfers using SoloHead sparse selection and a gather-scatter I/O engine. Implemented on SGLang, PulseInfer improves decode throughput by up to 4.7x over SGLang and 2.6x over the best existing offloading baseline, while reducing TPOT by up to 76% and preserving near-lossless accuracy.

Four-bit post-training quantization can reduce the memory demands of large language models, but preserving accuracy under strict MXFP4 W4A4 requires coordinating several design choices. Coordinate transforms change block-encoding errors, which in turn affect the residuals propagated through the network. The useful algorithmic decomposition is therefore not fully known before search. LLM-driven program evolution offers a way to explore these choices, but performance scores alone do not explain which design should change next. We introduce QuantForge, a PTQ discovery system that records competing explanations, selects controls that distinguish them, and checks that successor code implements the resulting conclusions. This residual compilation guides program revisions while retaining useful programs even when their original explanations are rejected. Remeasuring the revised program reveals the next error to address. This process discovers HiRes, a fixed MXFP4 quantizer that shapes coordinates, refines legal code assignments, and recovers errors along attention and MLP paths. Each stage acts on residuals measured after the preceding stage has executed. Across seven tasks, HiRes achieves the lowest seven-model Robust Fit (0.09300) and the lowest quantized Fit-7 at 32B. In matched-budget comparisons of LLM-driven program evolution, each with 240 evaluator calls, QuantForge reaches a held-out transfer target in six of eight runs, compared with three each for textual memory and reflection memory, and one for score-only evolution, despite evaluating fewer new programs. These results show that QuantForge improves the discovery of transferable PTQ algorithms by turning controlled evidence into subsequent program changes.

Long visual token sequences often account for a substantial fraction of the computational overhead in multimodal large language models~(MLLMs). Existing approaches reduce this cost by pruning redundant visual tokens, but permanently discard visual evidence that may become useful in subsequent layers. We instead ask whether all visual tokens can be preserved while reducing the cost of repeatedly evolving the representations through the Transformer. To answer this question, we perform low-rank interventions on visual-to-text information flow. We find that, after visual-to-text attention is blocked, restoring only a few directions recovers most of the lost accuracy, suggesting the relevant visual influence is concentrated in a low-dimensional subspace. We further observe strong predictability in layer-specific visual states: lightweight MLPs approximate them with high cosine similarity and low reconstruction error. Motivated by these findings, we propose $δ$-Vision, which replaces repeated Transformer evolution of visual tokens with lightweight low-rank adapters that construct layer-wise visual memories while preserving all visual tokens for text retrieval. Across image and video benchmarks, $δ$-Vision achieves higher accuracy than visual token pruning baselines at comparable or lower computation, while delivering competitive inference efficiency without discarding visual tokens.

Multimodal large language models (MLLMs) have demonstrated impressive performance in multimodal understanding, but processing large numbers of visual tokens results in high computational costs. While many methods have been proposed to reduce the number of visual tokens, most of them rely on heuristics and are prone to discarding substantial visual information during pruning, leading to degradation in model performance. In this work, by using a semantic erasure model, we derive a general mutual information coverage objective from task log-loss and propose MiCo, a training-free two-stage pruning method. MiCo first uses visual signals to select a representative candidate pool before visual tokens enter the language model, then performs task-aware subset selection within it. At each stage, suitable observable proxies instantiate the derived objective as a monotone submodular coverage function, which MiCo greedily optimizes under the token budget. MiCo is evaluated on diverse MLLMs ranging from 7B to 13B parameters across a broad range of image and video benchmarks spanning general visual reasoning, fine-grained OCR and grounding, hallucination detection, and long-video understanding. MiCo consistently achieves the best performance across nearly all evaluated models under all pruning ratios. On LLaVA-NEXT-13B, MiCo uses only 5.6% visual tokens, retains 97.5% of baseline performance, and achieves a 3.8-fold inference speedup. Our experiments demonstrate the effectiveness of MiCo and our mutual information coverage objective for visual token pruning.

Lookup-based memory has been a promising way to scale the parameters of large language models (LLMs). It retrieves learned representations of local token patterns, such as n-grams, instead of reconstructing them through successive layers of computation. However, existing designs such as Engram treat each retrieved embedding as a monolithic unit. Each embedding is stored in its own hashed slot and modulated by a single scalar gate. As a result, polysemous patterns cannot selectively read out the components of their memory that are relevant to the context. Moreover, parameters are shared only through hash collisions, which are largely unrelated to semantics. We propose FactorEngram, a factorized n-gram memory with basis-level contextual gating. FactorEngram retrieves sparsity-regularized coefficients over a dictionary of basis vectors shared across patterns, so related patterns can reuse common components. The same dictionary is also used for gating. The backbone hidden state is scored against each basis vector to gate the corresponding coefficient before reconstruction, which lets the context modulate each memory component individually. FactorEngram also covers both individual tokens and multi-token n-grams, and we systematically study where the memory branch should be inserted. On 340M- and 1B-parameter Transformer backbones, FactorEngram improves language modeling and downstream task performance. Ablation studies confirm the contribution of each component and identify insertion before the attention sublayer in the middle layers as an effective configuration.

Large language models (LLMs) achieve strong performance across many tasks but rely on dense multiply-accumulate (MAC) operations during inference, resulting in high energy cost. Spiking neural networks (SNNs) offer an event-driven alternative in which synaptic integration uses lightweight accumulation. However, spike-driven LLM inference remains difficult because outlier-heavy activations typically require long firing windows or auxiliary non-spiking paths. We propose QuantaSpike, a short-window spike-driven quantization framework for LLMs built around Logarithmic Ternary Integrate-and-Fire (LTIF) neurons. LTIF uses ternary events with power-of-two membrane-response quanta, improving the information represented by each firing step while retaining shift-ACC-compatible computation. QuantaSpike combines this neuron with group-adaptive gain and selective outlier admission: normal values use residual LTIF steps, whereas admitted outliers receive one additional onset spike before entering the same residual dynamics. Across OPT and Llama-2, QuantaSpike achieves state-of-the-art or competitive perplexity and zero-shot accuracy among spike-driven LLM quantization methods. It also transfers to newer dense LLMs, remaining close to the FP16 reference on Llama-3-8B and Qwen3-8B under the same four-step firing window. Analytical linear-energy projections show that QuantaSpike reduces the energy of one linear transformation by about $80.0%$ on OPT models and $67.1%$ on Llama-2 models relative to SpikeQuant, providing an accurate and energy-efficient spike-driven path for LLM inference.

Multimodal Large Language Models face significant efficiency challenges that stem from two distinct yet coupled sources: data redundancy and computational redundancy. While most methods focus on data redundancy by pruning visual tokens from the output of the visual encoder or computing redundancy in LLM decoders using blockwise importance, the finer-grained inter-layer representation shifts and the distribution differences within the layers themselves have not been fully explored. In this work, we comprehensively investigate this dual-level inefficiency. We posit that intermediate layer tokens from vision encoders should be considered for effective visual token pruning, as semantic focus shifts across layers, with middle-layer tokens capturing more detailed object-centric information that deeper layers may abstract away. Furthermore, we reveal the differential contributions of Attention and FFNs across distinct LLM decoder layers. Building upon these discoveries, we propose SPIDER, a training-free framework that integrates multi-layer \underline{S}emantic visual token \underline{P}run\underline{I}ng with an a\underline{D}aptive sub-lay\underline{ER} skipping mechanism. Experimental evaluations demonstrate that SPIDER consistently maintains strong performance across various MLLM architectures and reduction ratios. For instance, on LLaVA-NeXT-7B, SPIDER reduces FLOPs by $79%$ while maintaining 96$%$ of the baseline performance.

Multi-LoRA agent systems enable efficient role specialization by sharing a common backbone model. However, each agent repeatedly processes the growing shared trajectory and constructs its own KV cache, introducing substantial memory and computation redundancy in long-horizon tasks. Existing KV cache sharing methods reduce this repeated prefill, but they either require additional training or architectural constraints or retain substantial model computation. Moreover, direct cache reuse causes the current agent to rely on cache states generated by the previous agent's adapter, weakening the role-specific behavior encoded by its own LoRA. We present PReCache, a training-free KV cache sharing framework with two designs, namely PreLRShared and ReBaseShared, that share the base cache computed using the pretrained weights and precompute a compact agent-specific low-rank (LR) cache. To remove repeated prefill, PreLRShared precomputes each agent's LR cache when the shared context is first processed, allowing the current agent to use its own LR cache without reprocessing context processed by previous agents. To improve sharing accuracy, ReBaseShared reconstructs the shared base cache from adapter-free hidden states, reducing the remaining error caused by the previous agent's adapted representation. To minimize its reconstruction cost, we propose two inference schemes tailored to single-stream inference and concurrent serving, performing the same reconstruction after each agent's turn or alongside its execution, respectively. Across multiple models and agent benchmarks, PreLRShared achieves up to a 3.1x TTFT speedup and a 2.3x improvement in per-request throughput over inference without KV cache sharing. ReBaseShared best preserves accuracy overall among the evaluated cache-sharing methods, with an average drop of only 1.1 points relative to inference without cache sharing.

Reasoning vision-language models (VLMs) process images and videos as long sequences of visual tokens, making inference expensive. Training-free token pruning reduces this cost, but aggressive compression can sharply degrade performance, often attributed to irreversible loss of task-relevant visual information. We show that this explanation is incomplete. In a fixed-context Pass@K analysis, repeated sampling from the same pruned visual representation recovers many examples missed by greedy decoding, indicating that useful visual evidence can remain accessible but be used unreliably. We call this the representation-utilization gap. Motivated by this observation, we introduce SCOPD, a sparse-context on-policy self-distillation framework in which a student generates reasoning trajectories from pruned visual tokens while a privileged full-context teacher supervises the same on-policy prefixes. SCOPD requires no ground-truth responses, architectural changes, or additional inference-time computation. We further introduce SCOPD+, which uses a small visual-budget intervention to identify visually sensitive response positions and selectively distill them. At 10% visual-token retention, the Vanilla model retains 86.37% of its unpruned performance across 13 benchmarks. SCOPD raises this to 90.49%, while SCOPD+ further improves it to 92.43%. Across token budgets, benchmarks, and pruning operators, our results show that efficient reasoning depends not only on which visual information survives pruning, but also on how reliably the model learns to use it.

Between them, the members of a research group or a circle of friends own several consumer computers, none large enough to run a capable large language model. Existing systems pool such capacity across open swarms anyone may join, which a group admitting only trusted machines cannot use. Bounding membership removes what they depend on: a swarm holds each part of the model on several peers and routes around a slow one. A bounded session must use every device it admits. Its pipeline advances at the pace of whichever device received a share it cannot serve quickly, so the division has to be right before serving begins. We propose Kafila, whose protocol assembles a ring from behind NATs, preferring direct paths and relaying where traversal fails, while its planner measures each device's memory bandwidth, capacity and reachability, divides the model exactly for a fixed ring order, and places the head, which holds the embedding and output projection, together with that division rather than beforehand. On machines with different capabilities across three fleets, from a shared LAN to five devices spanning two continents, Kafila shortens the slowest pipeline stage by up to $5.2\times$ against the even split of pipeline parallelism, as in GPipe, and up to $3\times$ against the memory-proportional split of personal-device inference, as in exo, keeps 75 to 87 per cent of the committed hardware doing work where those divisions fall below half, and serves a model no uniform split can place on the fleet at all. What that is worth to a user depends on how much of a token is computation rather than network. Where the members share a network the same division returns $1.56\times$ the throughput of a uniform split and $1.25\times$ of a memory-proportional one, and under four concurrent users that lead compounds to $3.2\times$ rather than fading, each user served at almost the rate of one.

Sliding-window KV inference refers to processing a sequence incrementally while retaining only a fixed-size cache of recent key and value states. It can be applied to pretrained causal transformers at inference time without additional training, while its KV-cache memory remains fixed as more tokens are processed. Because cached states are computed in the context of earlier tokens, they may carry information from beyond the current window and transmit it to later states. This study presents a series of experiments using five open-weight models spanning Qwen, Llama, Mistral, and Muse Glimmer. We investigate whether information originating outside the immediate context window can persist through a rolling KV cache and remain useful for retrieval. Initial results show that retaining previously computed states improves retrieval across the models tested compared with recomputing the final fixed window from raw tokens. We then measure how far this effect extends and find that Muse Glimmer and Mistral 7B show the strongest latent information relay: they can recover information even after the relevant source tokens have left the cache. Both models incorporate sliding-window attention in their published architectures, an association that motivates testing whether training with sliding windows promotes more reliable information retention.

Mixture-of-Experts (MoE) models have emerged as a revolutionary method to scale Transformer models. However, traditional MoE architecture still suffers from inefficiency since a large number of experts are unnecessarily activated. Existing approaches for reducing the number of activated experts often overlook the historical performance of each expert. In this paper, we propose EAT, a novel method called Expert Account Tracker (EAT), which utilizes history-awareness metrics and adaptive thresholding to dynamically select the most important experts, thereby reducing the activated expert number while effectively maintaining the model performance. Experiments show that EAT outperforms the existing baseline Top-P method across multiple models and datasets, achieving over 25% an average reduction compared to the vanilla method in the number of activated experts and performing better token generation speed compared to the baseline. Furthermore, the performance of pruned models can be efficiently recovered via OPD using only 9K data. Additionally, through ablation studies, we find that excessively reducing the number of activated experts can significantly harm model performance, and the importance of experts varies across layers, with higher-level experts being generally more critical.

Multimodal Large Language Models (MLLMs) have demonstrated remarkable reasoning capabilities across vision and language tasks. However, their massive computational and memory demands hinder real-world deployment. While recent efforts reduce costs by employing lightweight language backbones, existing paradigms remain computation-dense due to their static sparsity and depth allocation, which cannot adapt to the semantic complexity of each token. To this end, we propose MoR-MLLM, a computation-sparse MLLM based on the recent Mixture-of-Recursions (MoR) framework. MoR-MLLM introduces adaptive per-token recursion, allowing the model to dynamically adjust its recursive depth and allocate more computation to visually or linguistically challenging tokens while skipping redundant operations for simpler ones. To stabilize the training of recursive sparsity in multimodal settings, we further design a three-stage MoR-Tuning strategy and an entropy-regularized loss to encourage diverse routing distributions. Extensive experiments show that compared with recent advanced tiny MLLMs, our proposed MoR-MLLM can greatly reduce the training memory and computation complexity while retaining high performance on various vision-language tasks.

Cross-context KV-cache reuse predicts a shared segment's keys and values under a new prefix instead of recomputing them, and has been reported to do so without quality loss. We find otherwise, and identify two problems. (1) A hidden cost: on MMLU and GSM8K, reuse costs substantial accuracy. (2) A decision at the wrong unit: no rule for deciding whether to reuse a cache removes that cost. What does help is choosing which parts of the cache to recompute, and the value of choosing well falls as the unit of choice grows: informed selection removes 49.5% of the cache error beyond chance at single rows (one token's keys and values), 10.6% at 64-token chunks, and nothing at the level of whole calls. Budgeted Cache Repair (BCR) acts at the unit where selection still pays. It drafts two tokens from the assembled cache, ranks cache rows by the attention those tokens pay them, and recomputes a fixed budget of rows exactly, in one of three layouts. The cost is paid rather than predicted away, and the draft that fails as a gate succeeds as a selector. BCR restores GSM8K to dense-prefill accuracy while still serving most calls from cache, and its best layout outperforms every reuse baseline's mean in the reference grid. The draft also beats a coin-flip selector at the same budget - a control prior evaluations lack.

Connectionist temporal classification (CTC) naturally supports offline and streaming speech recognition with utterance-level supervision, but conventional implementations materialize frame-by-vocabulary activations in memory, making CTC training with native LLM vocabularies prohibitively memory-intensive. A key observation is that every valid CTC alignment uses only target tokens and blank, and their union across a batch typically forms a small subset of the full vocabulary. We introduce Pruned CTC, which restricts alignment computation to this subset while retaining full-vocabulary normalization. We prove that this vocabulary reduction is exactly equivalent to full-vocabulary CTC in loss and gradients. Head-and-loss activation memory no longer scales linearly with vocabulary size. We further apply finite-beam alignment pruning. Building on Pruned CTC, we develop LLM-CTC, which adapts pretrained LLMs for non-autoregressive ASR while retaining causal attention and native vocabularies, and extend it to bounded-history streaming, avoiding chunk-level speech--text alignments. Experiments show that, with Zipformer-M encoder and 180K vocabulary, Pruned CTC reduces full-step memory by 5.1$\times$ with only 17% step-time overhead. Across three corpora, it matches standard CTC accuracy. On GigaSpeech, across six Qwen3 model sizes from 0.6B to 32B, LLM-CTC remains within 7% relative WER of LLM-CE with 7 to 10$\times$ faster recognition; when fine-tuning Qwen3-ASR for bounded-history streaming, LLM-CTC remains within 3% relative WER of matched offline models on the test set. Together, these results establish Pruned CTC as a scalable sequence objective for native-vocabulary LLM ASR across offline and streaming settings.

Low-precision Transformer systems increasingly quantize attention matrix multiplications, while softmax often remains at higher precision. During pretraining, an approximate softmax changes the gradients that train the model as well as its forward computation. We study this interaction with K-interval attention, which approximates the exponential using K+1 grid values. We vary per-row grid calibration, interpolation versus hard rounding, and the placement of a straight-through surrogate relative to normalization. We derive the corresponding backward rules, including calibration derivatives, and compare these choices in pretraining experiments matched on model, data, and optimizer. Detaching the row extrema leaves the forward computation unchanged but produces a delayed increase in validation loss. With hard rounding at K=4, min-max calibration and a pre-normalization surrogate incur a large loss gap; changing either choice substantially reduces it. At 124M parameters and 2.5B training tokens, fixed-window calibration with a post-normalization surrogate yields a validation loss gap of +0.019 nats relative to softmax at K=4, and with a pre-normalization surrogate yields +0.004 nats at K=16.

Hybrid-attention models replace most softmax attention layers with linear attention, reducing KV-cache growth and enabling larger decode batches where recurrent-state access becomes a major bottleneck. ReplaySSM amortizes state updates by buffering keys and values, but each new query still requires a full-state read even though the state remains unchanged between state updates. We observe that low-rank state-weighted query approximation accurately preserves state-read outputs. Although future queries are unknown, the basis vectors used to approximate them can be fixed offline. Based on this observation, we introduce SketchSSM, which preserves full-state updates while approximating reads. At each state update, SketchSSM reads the full state once to precompute outputs for these basis vectors, storing them in a compact sketch. Each subsequent decode step combines the sketch vectors with query-dependent coefficients to reconstruct the output without a full-state read. Across four Mamba-2-, GDN-, and KDA-based models, SketchSSM at a mean sketch rank of 8 reduces state-access traffic by approximately 10$\times$ while matching the average accuracy of the FP32 full-state baseline across four decode benchmarks, and preserves recall on four RULER retrieval tasks. At this rank on one NVIDIA B300, linear-attention kernel speedups over the Standard vLLM baseline reach 7.30$\times$, 5.02$\times$, and 5.24$\times$ for Mamba-2, GDN, and KDA, respectively, with up to 2.77$\times$ higher decode throughput on Nemotron 3 Super.

Training-free KV-cache compression methods such as SnapKV, H2O, and PyramidKV evict tokens at the end of prefill, aiming to preserve the attention mass that future queries are expected to use -optimizing what to keep. We show that this objective fails in two distinct ways. (1) Compensation: restoring the evicted attention mass can recover the attention-level target without recovering task quality. (2) Selection: covering more of the true decode-query mass can hurt quality when the recovered mass is fragmented rather than concentrated in coherent spans. These failures share a common cause: eviction occurs before the queries that determine the answer trajectory exist. We propose Draft-Guided Eviction (DGE), which defers eviction until after drafting the first k=2 answer tokens using the full cache - just one decode step beyond prefill. Because the draft is generated from the answer's own prefix, no cache entries are discarded before this trajectory signal becomes available. The per-head cache budget remains unchanged, and DGE can be applied directly to SnapKV, PyramidKV, H2O, and StreamingLLM without modifying their eviction scores. Unlike extra-pass methods, DGE changes when eviction occurs rather than what cache entries are selected. Extensive experiments demonstrate that DGE outperforms prior methods at every evaluated budget on five of six instruct-tuned backbones, achieving 44.2 on LongBench, nearly matching FullKV at 44.3. The timing-only control DGE-W achieves the same score, demonstrating that the gain comes from when eviction occurs rather than what is selected - an effect we term trajectory anchoring.

Cross-request KV caching reduces the prefill cost of Retrieval-Augmented Generation (RAG), but conventional prefix caching severely limits cache reuse across requests. Position-Independent Caching (PIC) removes this constraint by reusing independent chunks, but their KV states miss cross-chunk interactions. Existing methods selectively recompute token states to recover these missing interactions, but primarily allocate the recomputation budget to selecting which states to recompute, while fixing the recomputation context to the full causal prefix. We introduce RelaxKV, which formulates selective cache repair as a joint allocation problem over repair targets and recomputation context. Guided by the user query, RelaxKV identifies layer-specific repair targets and restricts their recomputation to a query-relevant context, reducing attention computation. Across four decoder models, RelaxKV at a 15% anchor ratio improves aggregate LongBench performance over ProphetKV on all models. On Qwen3-14B, RelaxKV provides a stronger quality-TTFT trade-off than ProphetKV across a 5%-30% anchor-ratio sweep, and achieves the best selective results on RULER-MV and LV-Eval at 16K and 32K context lengths. Controlled ablations further demonstrate the importance of recomputation context selection.

Genome foundation models are growing into sparse mixture-of-experts (MoE) networks whose expert weights no longer fit on the machines that hold the sequences, yet sending a private genome to rented accelerators exposes it: we show that a single server hosting one expert recovers the input nucleotides with 99.8% top-1 accuracy. We present CipherGenome, a protocol that keeps the embedding, attention and router of a 15.1B-parameter MoE genome model on a trusted thin client and outsources every expert projection, 95.8% of the parameters, to untrusted and possibly colluding GPU servers under module-LWE encryption. The design exploits three structural facts: expert layers are linear between two SwiGLU gates, expert weights are public, and GPU integer tensor cores can evaluate a ciphertext-weight product exactly modulo $2^{48}$ in a single GEMM. The client evaluates the nonlinearity exactly and re-encrypts with fresh secrets, so no polynomial approximation or bootstrapping is ever needed. On 72 windows from 12 bacterial genomes, encryption adds $2.54 \times 10^{-4}$ nats per token of KL divergence (95% CI upper bound $3.95 \times 10^{-4}$), below a pre-registered non-inferiority margin and indistinguishable from bf16 inference, while the same inversion attack falls to chance level. A reusable public hint cuts end-to-end latency by 3.54 times, wire compression reduces traffic 6.8 times, per-layer padding reduces routing leakage from 54.9% to 8.9% accuracy, and HE-compatible int4 experts remain non-inferior to their plaintext counterparts. Per expert and token, the server-side cost is more than six orders of magnitude below a CKKS baseline.

Genome foundation models are most useful where sequences are generated, yet the largest models need datacenter accelerators and a place to send private DNA. We ask whether a 15-billion-parameter genome mixture-of-experts (MoE) model can instead run on volunteers' web browsers, with the experts spread across many untrusted devices, without changing its predictions and without revealing the sequence to any single device. We build a system in which a trusted coordinator runs attention and routing while browser workers run every expert feed-forward network through hand-written WebGPU kernels, and we protect the expert inputs with real-valued Lagrange coded computing: each worker receives only a Gaussian-padded share, computes the expert's linear maps, and the coordinator decodes from any two of three workers. On GenomeOcean-MoE (8 experts, top-2 routing, 24 layers), the browser path matches native llama.cpp at every quantization level, the distributed path stays at the BF16 numerical noise floor (KL 0.0036 nats per token), and an unfitted latency model predicts decode time within 0.74% (median) under emulated wide-area links. We first show that plaintext expert inputs are not private: a probe recovers the token from a single vector at every depth, and one worker can identify the source genome from 300 unordered tokens with 92% accuracy. With coded experts, an adaptive attacker trained on shares falls to the most-frequent-token baseline, one worker's information about each token is bounded below one bit per forward pass, and the fidelity cost stays below the BF16 noise floor; in Chrome, coded decoding runs at 220 to 376 ms per token, depending on how much of the routing is hidden, and continues without replicas when a worker fails.

Policies with memory can learn along two backward paths: through the physical states their actions produce and through the representations they store. Transformer-XL and truncated backpropagation through time cut the second path at stored history while keeping its values. We ask when this cut matters. Holding the forward computation fixed and varying only derivative edges, we measure parameter gradients, the updates the optimizer applies, and continued training in a Transformer vessel-trajectory model and a quadrotor tracking policy. In the vessel model, detaching the key-value cache shrank the gradient to about a tenth of its norm, with little rotation, when gradients flowed through all earlier physical states, but barely changed it under one-step physical credit. In this strongly clipped regime the optimizer, not the gradient, set how far updates differed: global-norm clipping removed most of the gradient difference between memory-cut graphs, whereas AdamW turned a 2% gradient difference between two placements of the cut into update differences of up to 31% at the step where the placement was switched. In a quadrotor trained from initialization with 0.20 m/s velocity noise, removing memory raised tracking error by 43% and cutting memory gradients raised it by 32%; at low noise the cut's mean cost exceeded the value of memory. Two-step truncation segments gave no measurable gain, although with hidden velocity a two-step window captured most of the value of memory; eight-step segments removed half to three quarters of the cost. Switching the cut on only for the last fifth of training understated its cost about threefold at 0.20-0.30 m/s, but not at low noise or with hidden velocity. These results suggest measuring the cost of a memory cut by training with it from initialization, and comparing backward graphs by the updates the optimizer applies rather than by raw gradients.

Quantization errors in video diffusion transformers can be amplified or attenuated by subsequent denoising updates, making local reconstruction error an incomplete predictor of final impact. We introduce PulseQuant, a 4-bit post-training quantization method that combines trajectory sensitivity with activation geometry to guide offline calibration. Isolated block--step interventions estimate propagation risk, which prioritizes sensitive trajectory states during row-radius selection. With these radii fixed, response-subspace correction uses neighboring-code edits to reduce residual components along dominant activation directions. Both stages preserve the original 4-bit weight representation. Controlled interventions show that short-horizon propagated error predicts final latent error more reliably than immediate block-output error, supporting calibration beyond local reconstruction objectives. Evaluations on Wan models, Self Forcing, and MiniMax-H3 demonstrate improvements in key consistency and dense-reference metrics while remaining competitive on other attributes across model scales and generation paradigms.