生成模型与LLM推理优化

Reasoning models spend different amounts of useful computation across instances, but it remains unclear when a learned stopping rule improves over simple confidence or convergence thresholds. We study this question with LearnStop, a hidden-state-free checkpoint stopper for reasoning language models. At fixed budget checkpoints, LearnStop probes a short answer from the current reasoning prefix and predicts prefix correctness from online features such as answer confidence, entropy, prefix vote share, answer stability, and backtracking-marker density. Across 18 task-model settings spanning GSM8K, MATH-500, MMLU-Pro, AIME-90, GPQA, Qwen3, and DeepSeek-R1 distillations, the answer is task-dependent. On free-form math, learned multi-feature stopping improves the fixed-budget frontier and often beats scalar exits: on GSM8K with Qwen3-32B, the empirical frontier reaches a post-hoc peak adapt gain of +0.157, validation-selected operating points preserve positive gains, and the paired gain over the strongest scalar baseline is +0.028. On multiple-choice and very hard settings, scalar confidence, entropy, or stability rules are competitive or stronger. We therefore frame learned stopping not as a universal replacement for scalar exits, but as a tool whose value depends on trajectory structure. We further provide validation-selected operating points, paired bootstrap tests, finite-grid lost-correct risk calibration, cost accounting under KV-fork, prefix-cache, and black-box regimes, H100 serving profiles, checkpoint-schedule sweeps, transfer analyses, and robustness checks. The main practical finding is that learned stopping is useful when many questions become correct before full budget but do not exhibit a single reliable scalar stopping signal; its benefits largely disappear when confidence or answer convergence already solves the stopping problem.

Scaling fault tolerant quantum computing is increasingly constrained by the limited bandwidth and power budget across the 4 K to room temperature (RT) interface. We present CryoZip, a cross stack cryogenic compression framework that cooperates with a lightweight cryogenic quantum error correction (QEC) predecoder to reduce 4 K to RT syndrome transmission under realistic, circuit level noise. CryoZip targets sparse syndrome vectors with a sliding window compression architecture sized under strict decoding latency constraints to maximize energy efficiency. We implement and evaluate the design in 22 nm FDSOI characterized at 4 K, using vector based power, performance, and area analysis to obtain realistic hardware data. CryoZip achieves up to 48x compression, 1.8x higher than state of the art compressors, across various QEC codes while delivering 4 to 26x energy savings. When paired with a QEC predecoder, it yields over 14,238x bandwidth reduction, while energy savings rise to 42x when accounting for realistic QEC interface overheads.

Hierarchical Global Attention (HGA) is a drop-in replacement for dense causal attention in pretrained long-context transformers. HGA preserves the original checkpoint parameters: the pretrained $W_Q$, $W_K$, $W_V$, and $W_O$ projections remain unchanged, no calibration parameters are introduced, and no retraining is required. Applied to Qwen3-30B-A3B-Instruct-2507-FP8 on a single RTX~5090 (32GB), the patched model runs out of the box at a 64K-token context, where token-level K/V storage is not feasible on this hardware. Unlike previous sparse-attention methods, HGA performs hierarchical two-level routing. It first retrieves relevant chunks using compact RoPE-aware summaries and then refines the selection by routing only the most relevant groups before performing exact token-level attention. This hierarchical retrieval significantly reduces the number of fetched tokens while preserving exact attention over the retrieved token set, making RAM- and NVMe-backed storage practical. The full historical token K/V resides in host RAM or NVMe storage, while only a small routed working set is transferred to GPU memory during attention. Consequently, GPU memory consumption depends primarily on model weights and the routed working set rather than on the total context length. Across all tested context lengths (4K - 64K tokens), routed attention remains within approximately $0.01$--$0.02$ nats of dense attention while the sparsity used is just about 3%. These results suggest that the approximation introduced by hierarchical routing is small, and that the remaining quality gap is likely dominated by long-context positional encoding rather than by the routing algorithm itself.

Diffusion Transformers (DiTs) have significantly advanced audio-driven portrait animation, but their high computational cost leads to substantial inference latency. Although training-free diffusion caching accelerates inference significant, existing methods are primarily developed for text-conditioned generation and overlook the spatial and modality imbalances inherent in audio-driven portrait animation. In this paper, we propose SyncCache, a training-free caching acceleration method tailored for DiT-based portrait animation that explicitly exploits asymmetric dynamics. Specifically, high-frequency dynamics driven by audio conditions and concentrated in human regions are more challenging and critical to cache and reuse than the low-frequency visual background in portrait animation. First, we introduce Spatially-Asymmetric Probing to prioritize error sensitivity in dynamic human region. Second, through Modality-Decoupled Caching, we bypass heavy DiT block by reusing stable inter-block residuals, while continuously recomputing lightweight audio blocks to preserve precise lip synchronization. Furthermore, we introduce a cache ratio to control cache capacity and formulate memory-adaptive cache selection as an offline dynamic programming problem without online overhead. Extensive experiments demonstrate that SyncCache achieves superior speed-quality trade-offs, delivering up to 4.12x acceleration on HunyuanVideo-Avatar and 3.75x on Wan-S2V with near-lossless visual fidelity and precise audio alignment.

Coding agents are rapidly becoming a major application of agentic LLMs, but serving them efficiently remains challenging. Progress on this challenge requires understanding real workload patterns, yet the data needed for such analysis is largely absent. Existing public traces and benchmarks do not capture real, day-to-day coding-agent usage across multiple agents and model families for serving-system analysis. To help fill this gap, we collect and release a trace of roughly 4,300 coding-agent sessions, containing about 350,000 LLM steps and 430,000 tool calls from our own day-to-day use of Claude Code and Codex. Our analysis shows that coding-agent workloads feature long autonomous loops, long contexts with short outputs, diverse and heavily-tailed tool calls, and high but imperfect prefix cache hit rates. These findings point to concrete opportunities for optimizing serving, including lower-overhead tool calling, append-length-aware prefill, semantic-aware tool-latency prediction, and improved KV-cache management around human-paced gaps. We release the dataset, trace collection pipeline, and analysis code at https://github.com/uw-syfi/TraceLab.git the project website is https://tracelab.cs.washington.edu.

The development of on-device large language models (LLMs) is driven by the need for privacy and fast response times. Energy-intensive data transfer on mobile devices makes Processing-in-Memory (PIM) an effective solution. Due to stringent DRAM cost constraints, limited physical footprint on circuit boards, and the interaction between applications and LLMs, it is imperative for the CPU and PIM to operate concurrently within a shared memory space. However, challenges such as bank conflicts and bus congestion can arise, potentially diminishing the performance and energy benefits of PIM. To address this challenge, we introduce COSM, a cooperative scheduling framework designed to facilitate the concurrent operation of PIM and CPU tasks on mobile platforms. Our key innovations include: 1) a low-interference PIM control interface that generates the maximum number of PIM commands without disrupting CPU memory accesses; 2) an idleness-aware scheduling method that integrates PIM commands into available idle time windows within the CPU's access sequence. COSM not only hides PIM execution latency from the CPU, but also overlaps PIM execution with data transfer. Experiments on concurrent execution of LLMs and mobile workloads, including mobile applications and compute-intensive kernels, demonstrate that COSM improves PIM throughput by up to 2.8x compared to the baseline scheduling method with less than 2.0% CPU performance loss.

DiT video generation is latency-intensive due to iterative full-frame denoising, while prior cloud-edge methods largely rely on static inter-step decoupling and cannot leverage inter-frame similarity or adapt to system dynamics. We propose EcoVideo, an entropy-orchestrated framework for dynamic inter-frame decoupling: early-stage self-attention entropy provides a training-free estimate of frame-wise information density for frame selection; a cloud large model denoises sparse high-entropy keyframes; and an edge lightweight model reconstructs the remaining frames via motion-aware interpolation with refinement for temporal stability. EcoVideo further adapts the keyframe budget and edge refinement depth to real-time bandwidth and compute availability, optimizing end-to-end latency under constraints. Experiments on representative DiT video generators show improved quality--efficiency trade-offs and up to 2.9x end-to-end speedup in low-bandwidth, compute-limited edge settings. Code is available at https://github.com/IF-LAB-PKU/EcoVideo.

As LLM inference becomes a major cloud workload, its growing energy footprint makes cluster-wide energy optimization increasingly important. Serverless LLM serving helps platforms absorb traffic volatility by elastically sharing GPU resources across models, but this sharing also makes energy optimization difficult. Multiple co-resident models run under one device-wide operating point, while their resource demands and latency slack change across execution phases and load conditions. As a result, minimizing energy requires coordinated scheduling across request placement, runtime resource adaptation, and workload consolidation. We present Festina, a profiling-guided, power-aware control plane to minimize cluster-wide energy for serverless LLM serving. Unlike common global-local schedulers that focus on throughput or tail latency, Festina makes energy-first decisions by jointly coordinating request placement, SM partitioning, and GPU operating points under TTFT/TBT SLOs. In our system, a lightweight global scheduler performs fast, SLO-safe, energy-aware placement using constant-time lookups from offline profiles and GPU state summaries. On each GPU, a phase-aware local scheduler continuously adapts task batching and compute resources to minimize power consumption. Festina further performs energy-aware workload consolidation to reduce GPUs' static power consumption via SLO-aware migration. Comparison with four SOTA LLM serving systems and one DVFS-augmented system demonstrates that Festina reduces energy consumption by up to 56% while maintaining parity in SLO attainment (within a 2% margin)

Dynamic sparse attention (DSA) accelerates long-context LLM decoding by attending to only the top-K KV blocks relevant to each query, but it introduces a serialized selection-to-attention dependency that emerges as a new latency bottleneck. We present PRR, a speculate-reuse-repair runtime that exploits temporal locality in DSA selections to predict likely blocks, speculate the attention over them while selection is in flight, and incrementally repair missed blocks once the true selected set is known. PRR uses a lightweight EMA-based predictor, a profiling-guided speculation budget that keeps speculative work off the critical path, and a FlashAttention-based repair kernel that folds missed blocks into the partial attention state using online-softmax statistics. Across long-context benchmarks and representative DSA methods, PRR reduces per-token decoding latency by up to 40% while preserving downstream task accuracy. Github: https://github.com/Tianyu9748/Incremental_FlashAttention

Convolutional Spiking Neural Networks (SNN) offer the potential for highly energy-efficient vision processing by exploiting sparse, event-driven computation. However, existing SNN accelerators underutilize the inherent parallelism of convolutional layers and lack the flexibility to accommodate varying memory demands and input sparsity across layers. This paper presents Mega, a digital architecture for convolutional SNNs that addresses these limitations through three key contributions: (1) highly parallel acceleration of $3 \times 3$ convolutions, (2) a unified data memory for spikes, neuron states, and weights, and (3) efficient spike map processing with low-overhead spike detection. Fabricated in GlobalFoundries 22 nm FDSOI technology, Mega achieves an energy efficiency of 0.375 pJ/SOP, improving the state of the art by $4\times$.

Discrete action tokenization provides a compact interface for autoregressive VLA policies, but accurately recovering continuous robot actions from discrete codes remains challenging. Existing tokenizers typically map each discrete code to a fixed continuous action prototype, ignoring the robot's current proprioceptive state. This limitation is particularly pronounced in manipulation, where the same action token may require different continuous controls under different joint configurations, object poses, and contact conditions. We therefore propose SA-VLA, a state-aware action tokenizer that conditions action decoding on robot state. We study two state-injection mechanisms for VQ-based action tokenization: cross-attention between state and action features, and a lightweight state adapter that predicts action-wise modulation factors for state-conditioned action modulation and reconstruction. The adapter formulation expands the effective support of a finite codebook by allowing each discrete token to represent a family of state-dependent continuous actions, while preserving the efficiency and compatibility of discrete action modeling. Integrated into an LLM-based VLA policy, SA-VLA supports both autoregressive and parallel action-token decoding with minimal changes to the model interface. On 12 RoboTwin manipulation tasks, SA-VLA improves the average success rate from 0.29 to 0.56 over the strongest tokenizer baseline. In zero-shot sim-to-real experiments on three real-world tasks, it further improves average success from 0.15 to 0.33 over the strongest tokenizer baseline. These results demonstrate that state-conditioned action decoding is a simple and effective mechanism for reducing the compression gap in discrete VLA policies.

Agentic search equips large language models with dynamic retrieval abilities, but existing reinforcement learning methods remain limited by reward sparsity in knowledge boundary calibration — deciding when to trust parametric memory, when to rely on retrieved evidence, and when to abstain. Binary rewards can penalize undesirable outcomes, but provide little guidance on the reasoning process required to make calibrated decisions across different knowledge states. To address this, we propose KbSD (Knowledge boundary Self-Distillation), a framework that tackles this limitation through dense token-level supervision, outcome-level sparse rewards, and quadrant-adaptive optimization. KbSD constructs a hint-augmented teacher, architecturally identical to the student, that receives explicit knowledge boundary signals — including parametric certainty, retrieval quality, and ground-truth answers — to generate calibrated reasoning demonstrations. This information-asymmetric self-distillation enables dense supervision without requiring a larger external model. To further account for the heterogeneous reasoning distributions across knowledge states, we introduce a quadrant-adaptive distillation objective: reverse KL for concentrated integration, forward KL for diverse refusal, and Pareto-optimal bidirectional KL for asymmetric quadrants requiring both precision and coverage. Experiments on multiple benchmarks show that KbSD consistently improves both task accuracy and hallucination mitigation over strong baselines, with the largest gains appearing in the challenging quadrants where sparse rewards are least informative.

The quadratic computational cost of traditional attention mechanisms poses a major bottleneck to the scalability and practical deployment of large language models (LLMs), particularly in long-context scenarios. To improve efficiency, existing approaches often enforce rigid structural constraints such as local attention windows. However, these strategies typically lead to substantial performance degradation on tasks requiring precise long-range recall. In this work, we propose MATCH, a scalable and efficient framework that augments sparsified attention mechanisms with dynamically integrated in-context information through an efficient retrieval system. Empirical results show that MATCH significantly improves the performance of sparse-attention models on both synthetic and real-world natural-language tasks. These findings highlight the versatility of MATCH as a general approach for enhancing in-context retrieval capabilities while maintaining the efficiency benefits of sparse attention architectures.

Speculative decoding accelerates language model inference by using a fast drafter to propose candidate tokens that are then verified by a larger target model. Existing theory largely studies the stochastic, distribution-preserving setting, where the goal is to exactly sample from the target distribution. In contrast, many practical systems use greedy decoding, relaxed acceptance rules, or tree-based candidate sets, where success is governed by local ranking and threshold events rather than exact distributional equality. We develop a theory for these regimes. We identify that many common acceptance criteria have rejection regions that can be characterized as lower level sets of the target distribution. For these, we characterize the exact KL divergence required for rejection yielding exact certificates and sharp margin-based bounds for strict greedy decoding, additive and multiplicative relaxed acceptance, top-(m) relaxed criteria, and entropy-thresholded acceptance. We then extend the framework to greedy tree decoding, deriving exact and margin-only certificates for when the target greedy token remains covered by the drafter's top-(m) candidates. Finally, we evaluate the resulting certificates on Qwen3 models, showing that relaxed and tree-based criteria substantially enlarge the region of certified acceptance, especially on decoding steps with low target model distribution margin. These results complement existing distribution-preserving analyses of speculative decoding by characterizing the deterministic local acceptance events common in practical inference systems.

Mixture-of-Experts (MoE) architectures enable language models to achieve unprecedented scale via sparse activation. However, their inference performance is often limited by data movement bottlenecks. Two coupled challenges exacerbate this limtation: (1) Importance-Agnostic Cost: Low-contribution experts incur nearly uniform memory and transfer costs, resulting in a low cost-to-benefit ratio and wasting critical bandwidth; (2) System-Level Imbalance: Multi-device deployments are universally bottlenecked by the slowest device, meaning that local reductions on one device may yield no improvement in end-to-end latency. We propose Cost-Aware Expert Execution (CAEE), a hardware-guided runtime framework that jointly optimizes for token-level expert importance and system-level execution cost. CAEE uses lightweight, calibrated cost models to estimate hardware overhead, selectively prunes low-importance, high-cost experts, and redistributes their contributions via a low-overhead compensation mechanism, avoiding extra data movement. Evaluations on the 671B DeepSeek-R1 model show that CAEE can reduce end-to-end inference latency by 8%-18% across diverse deployment settings, including expert offloading and on-device execution on multi-device systems, while maintaining a model accuracy drop of less than 1%.

Heterogeneous prefill-decode (PD) inference is now in production: prefill on cost-efficient or supply-available accelerators, decode on bandwidth-strong ones, and KV state crossing mixed interconnects in mixed numerical formats. Each deployment makes these decisions on its own. What is missing is the picture across configurations-which decisions must be made jointly at the PD boundary, and which can be made independently. We propose a design space organized along four design axes-accelerator, precision, interconnect, and KV residency and the workload regime (stage pressure) they respond to. We show that only a subset of interactions among these factors become binding constraints once PD inference becomes heterogeneous. These interactions surface through three recurring boundary decisions: compute placement, KV representation, and KV ownership. The resulting analysis yields concrete guidance. Precision policy belongs to runtime roles rather than to a single system-wide setting, because the same low-bit format relieves different bottlenecks on each side of the boundary. KV transfer engines move bytes rather than tensor semantics, making representation compatibility an explicit boundary concern whenever producer and consumer differ. The KV handoff also carries a lifecycle-reservation, release, and failure recovery-that spans prefill and decode and requires explicit ownership. Two further interactions remain open. Cross-vendor and interconnect-related claims are stated as design guidance grounded in industrial deployment observations and source-code inspection of the runtimes involved.

LLM inference comprises a compute-bound prefill phase and a memory-bound decode phase, and recent systems disaggregate them onto separate hardware. Yet today's datacenter GPUs rely on costly HBM whose bandwidth sits almost entirely idle during prefill. LLM serving across memory-heterogeneous accelerators (MemHA) pairs GDDR-based accelerators for prefill with HBM-based GPUs for decode, promising lower cost without sacrificing performance. Pushed to its most economical form, MemHA serving is inherently cross-vendor, since the best-suited chip for each phase may come from a different vendor. This breaks two assumptions that single-vendor disaggregation takes for granted — a KV format both ends consume natively, and a shared software stack. We present HMA-Serve, a MemHA-centric disaggregated serving system pairing GDDR-based accelerators for prefill with HBM-based GPUs for decode efficiently. HMA-Serve achieves this through (1) phase-wise quantization, applying vendor-native low precision for high-throughput prefill while keeping decode in high-precision BF16, (2) a compute-transfer pipeline that overlaps each layer's KV cache transfer with later-layer prefill to reduce time-to-first-token (TTFT), and (3) deferred dequantization, shipping raw quantized bytes and reconstructing them lazily on the decode GPU to reduce network bandwidth and HBM usage. Across four Qwen3 models (4B--32B) and three production traces, HMA-Serve delivers up to $3.2\times$ higher goodput than state-of-the-art memory-homogeneous methods and $4.8\times$ higher goodput-per-dollar, with no measurable loss on generation-quality benchmarks.

When does retention matter for memory-augmented LLM agents? We study this with TraceRetain, a lightweight framework for bounded external memory in frozen LLM agents that scores entries by interpretable features (success, age, access frequency, redundancy, specificity, similarity, downstream utility) and evicts the lowest-scoring ones at capacity. On clean ALFWorld with gpt-5-mini, external memory robustly improves over no memory across two seeds, but differences among bounded retention policies fall within Wilson 95% CIs: clean ALFWorld at T=100 to T=200 does not naturally exhibit the memory pollution retention is designed to address. Under a controlled noisy-write stress (75% synthetic distractors), unbounded memory and FIFO-K50 degrade on Precision@5 (20.2% to 12.4% and 15.8% to 3.8%) while TraceRetain-CEM is essentially unchanged (16.9% to 16.6%) and preserves 97/100 task success. The mechanism: unbounded memory has the highest mean similarity (0.87) but lowest precision, indicating failed distractors close to the query in embedding space. Held-out in-distribution evaluation shows memory-augmented policies solving 47 to 49 of 50 tasks vs. 39/50 for no memory. Bounded retention buys memory and step efficiency on saturated clean benchmarks at no task-success cost, and only differentiates from cache heuristics when streams contain noise.

Fog computing utilizes proximal computational resources for sensor data processing and actuation, and addresses the latency, network load, and privacy issues of cloud-centric Internet of Things. On the other hand, Large Language Models (LLMs) are a type of deep learning AI models, which are trained on enormous text data, that perform various natural language processing tasks such as translation, question answering, text summarization, and code generation. LLMs are generally cloud-centric, requiring abundant GPU memory and computing capabilities, again face the same issues that led to fog computing. This pushes the necessity for LLM support in the proximity on fog infrastructure, requiring LLM optimizations such as parameter-weight quantization, pruning, low-rank adaptation etc. Meanwhile, fog computing also gets benefit from LLM's ability for code generation, in the dynamic deployment of fog-based applications. The paper addresses how both fog computing and LLMs can be mutual beneficiaries, discussing the state-of-the-art and future research scope.

Block Diffusion Language Models (BD-LMs) improve diffusion-based text generation with KV caching and flexible-length generation. A natural next step is to extend them from Single-Block Diffusion (SingleBD) to Multi-Block Diffusion (MultiBD), where a running-set of consecutive blocks is decoded concurrently for inter-block parallelism. However, existing BD-LMs are mostly trained under teacher forcing, where the model observes only one noisy block conditioned on a clean prefix. While the recent diffusion forcing strategy introduces visibility among multiple noisy blocks, its training states still differ from MultiBD inference, where decoding operates on a bounded running-set with heterogeneous slot-wise noise patterns. To bridge this gap, we propose Multi-Block Diffusion Language Models (MBD-LMs), obtained by post-training BD-LMs with Multi-block Teacher Forcing (MultiTF). MultiTF integrates teacher forcing and diffusion forcing by training on bounded noise-groups conditioned on clean prefixes, with randomized noise-schedulers that better match MultiBD inference states. To make MultiBD practically executable, we further introduce an optimized decoding algorithm based on the Block Buffer mechanism that preserves prefix-cache reuse, keeps input shapes static, and translates increased decoding parallelism into wall-clock acceleration. Empirically, MBD-LLaDA2-Mini increases average Tokens Per Forward pass (TPF) from 3.47 to 6.19 and improves average accuracy from 79.95% to 81.03%; when combined with DMax, MBD-LLaDA2-Mini-DMax reaches an average TPF of 9.34 with only a 1.02% accuracy drop on math and code benchmarks.

We summarize our submission to Sub-Challenge 1: W4A4 Quantization for Inference (HiF4 / MXFP4) of the ICME 2026 Low-Bit-width Large-Model Quantization Challenge. The sub-challenge targets 4-bit weight and 4-bit activation inference on Wan-AI/Wan2.2-I2V-A14B under HiF4 or MXFP4 numerical formats. We adapt two complementary ideas from LLM quantization, MixQ-style mixed precision for sparse activation outliers and SmoothQuant-style per-channel smoothing, together with block-wise HiF4 packing for Wan2.2 feed-forward linear layers. Calibration on representative OpenS2V-5M batches identifies heavy-tailed activation channels; smoothing rebalances dynamic range before W4A4 rounding; and a dual-branch GEMM preserves outlier columns in higher precision while the bulk of channels use strict W4A4. On official VBench I2V metrics, our pipeline stays within 2-3.5 percent of FP16 on most quality axes and improves motion smoothness, outperforming a native HiFloat4 baseline that degrades roughly 5 percent relative to FP16 across all reported scores.

Multimodal model inference creates substantial energy demand with growing performance requirements. Within GPUs, power is autonomously managed by an on-board power management unit (PMU), which makes frequency boosting/throttling decisions. However, we find that these hardware-managed frequency decisions can cause significant power inefficiency. This work identifies three classes of power inefficiencies within modern multimodal inference serving: (1) inter-stage dependency stalls run at near maximum frequency despite being idle; (2) anti-correlation between auto-boost frequency and arithmetic intensity (A.I.) results in compute-bound phases (e.g., prefill) running at lower frequency and vice versa; and (3) thermal throttling degrades SM frequency and throughput. We propose Tri-serve, a software-based DVFS controller that jointly accounts for three classes of inefficiency — inter-stage Dependency stalls, the Arithmetic-intensity effect on frequency and power, and the Thermal-throttling effect of high A.I. phases — to deliver energy-efficient multimodal serving on commodity GPUs. We show that Tri-serve achieves 22% energy efficiency improvement with no latency or throughput impacts.

Modern LLM deployments routinely compress models and raise sampling temperature to reduce cost, latency, or repetition, yet safety evaluations usually treat these choices as fixed implementation details. This leaves a practical uncertainty: does a model that is safe at FP16 and greedy decoding remain safe after it is quantized and sampled stochastically, or do the two deployment knobs amplify one another? We study this question with a factorial evaluation of 9 instruction-tuned models from six families, 3 precisions (FP16, GPTQ INT8, AWQ INT4), and 6 temperatures ($T{=}0$ to $1.0$), yielding 161 configurations and $\approx$322k responses judged by a six-model safety ensemble. Contrary to the concern that low-bit deployment broadly erodes alignment, standard non-adversarial quantization is usually safety-neutral: INT4 keeps or lowers attack success for 7 of 9 models, with clear degradation concentrated in the weakest baseline model, SmolLM3-3B ($18.5%{\to}36.0%$). The larger risk comes from sampling: higher temperature sharply increases decision instability for vulnerable models, with DFR reaching 53.0% at $T{=}1.0$, even when average ASR changes modestly. Finally, the interaction is not a ``double penalty'': our Compound Degradation Index remains largely sub-additive ($-0.195$ to $+0.045$), indicating that quantization and temperature do not systematically compound. These results suggest a deployment rule of thumb: standard INT4/INT8 quantization can be reasonable for strongly aligned models, but safety claims at elevated temperature should report multi-sample stability, not only average attack success.

A stateless inference server (vLLM, SGLang, TensorRT-LLM) idles between requests while the accelerator waits; a stateful session reclaims that idle time. Speculative pre-positioning decodes the session forward to its next decision point with the target model's own forward pass and no draft model, moving the cross-request prefill and entry-decode off the critical path: the next request resumes from a pre-paid entry on its delta, or, when a confidence gate fires, is answered from a cached distribution in one near-constant vocabulary scan with no decode, at a cost only of energy and a rare, bounded false accept. The payoff is conditional on capability: a capable model fires the gate at near-full coverage and about 87% precision (a smaller one never clears it), returning the first token in about 1.0 ms versus the 39 ms decode a prefix cache still pays.

Large language models (LLMs) excel at complex tasks like question answering and summarization, thanks to their ability to handle long-context inputs. However, deploying LLMs is costly, not only due to the high computational demands of quadratic complexity of self-attention and auto-regressive generation, but also because of the significant memory overhead required for storing the key-value (KV) cache during inference. To reduce the memory cost, existing KV-cache eviction strategies leverage the sparsity in attention to selectively store a subset of tokens. While reducing the memory footprint, such approaches show a considerable drop in performance, especially in tasks that require long-context reasoning. We identify that the drop in performance is linked to a reduction in the coverage of unique tokens. Additionally, we theoretically show that reduced coverage limits the mutual information between inputs and outputs, thereby impairing predictive accuracy. To this end, we introduce K-VEC, a novel coverage-aware KV-cache eviction strategy that prioritizes token coverage while evicting tokens in the cache. K-VEC introduces a cross-head and a cross-layer coverage module to enhance token retention across attention heads and model layers, mitigating performance degradation caused by low coverage. Evaluated on 16 LongBench subsets, K-VEC exhibit up to 10.35 points improvement over the existing methods under the same eviction rate and memory constraint. Comprehensive evaluations validate the effectiveness of our approach and demonstrate its potential for efficient LLM deployment in resource-constrained settings.

Model quantization is essential for the efficient deployment of Large Language Models (LLMs), but introduces a critical vulnerability: Quantization-Conditioned Backdoor (QCB) attacks. In these attacks, malicious behaviors remain dormant in full-precision models and activate only after specific quantization distortions, bypassing standard security audits. To mitigate this, we introduce FlipGuard, a proactive defense framework that selectively perturbs model weights prior to quantization. By breaking the adversary's precise alignment between weight patterns and quantization boundaries, FlipGuard suppresses backdoor activation without requiring access to training data or trigger samples. We further propose the Defense Effectiveness Ratio (DER), a unified metric to jointly evaluate security gains, utility preservation, and computational cost. Extensive experiments across seven LLMs (including StarCoder and LLaMA-family models) and three quantization schemes (INT8, FP4, NF4) demonstrate that FlipGuard effectively neutralizes QCBs across three scenarios, i.e., vulnerable code generation, content injection, and over-refusal, achieving high security with negligible performance degradation.

Long-context language models often conflate two different goals: compressing history into an efficient state, and maintaining reliable long-term memory. Linear, recurrent, and sparse attention reduce the cost of processing long sequences, but they do not by themselves specify when a fact should be written, overwritten, protected from distractors, or discarded. We study memory-managed long-context attention, a research route that separates a fast recurrent or sparse backbone from explicit editable request-local memory slots and query-time sparse fallback. Across structured synthetic tasks, token/chunk/sequence bridges, generated natural language, and local frozen-model diagnostics, pure fixed-state or pure sparse methods fail some overwrite, version, anti-pollution, or no-write-signal cases, while a hybrid covers both routes. A small 2,097,152-token mechanism stress test reaches 50/50 pooled accuracy with 2-132 active chunks. A 2.74M-parameter minimal causal event-token model reaches 595/600 with lite write supervision, supporting proof of trainability rather than scale. A six-family frozen-hidden-state bridge reaches 1079/1080 controlled pointer accuracy, but it uses generator-provided integer key IDs and separately encoded canonical key strings; it is an oracle-metadata probe, not open-text entity resolution. Local non-leaderboard RULER 4K diagnostics remain close to full context, whereas a 33-record LongBench v1 16K subset shows that naive lexical selection is not general. The evidence separates three claims: controlled slot lifecycle is feasible, sparse fallback is needed when writes lack future-query signals, and learned open-domain selection remains the main architectural bottleneck. We do not claim a final generative architecture, global slot-trajectory convergence, or systems superiority.

Diffusion language models (DLMs) have recently emerged as a promising alternative to conventional autoregressive language models. By generating multiple tokens in parallel during each denoising step, they offer higher inference throughput while maintaining competitive quality. However, realizing these throughput gains while meeting latency SLOs in a serving system requires addressing challenges introduced by DLMs' unique characteristics. These include navigating the speed-quality tradeoff created by confidence-based denoising, choosing appropriate parallelization levels across model instances under fluctuating load, and coordinating approximate KV caching mechanisms that introduce non-uniform per-step costs. To address these challenges, we present DiLaServe, a cluster-level serving system for DLMs. DiLaServe enables deadline-aware scheduling and adaptive load control through confidence-threshold adjustment, and dynamically reconfigures the cluster by solving a quality-aware optimization problem, while explicitly modeling the step-level heterogeneity introduced by approximate KV caching. Across multiple benchmarks and real-world traces, DiLaServe improves SLO attainment by up to 56.6 percentage points and reduces end-to-end request latency by up to 46% while incurring less than 1% accuracy drop.

Long-context LLM inference faces a fundamental conflict: head-adaptive compression algorithms (e.g., Top-$p$ nucleus sampling) offer superior accuracy by dynamically fluctuating memory budgets, yet modern inference engines (e.g., vLLM) demand rigid, static memory patterns to leverage CUDA Graphs and PagedAttention. We resolve this ``Static-Dynamic'' mismatch with HARD-KV, a unified framework that that bridges dynamic selection with rigid system constraints. HARD-KV introduces a Cascade Cache hierarchy, managing the token lifecycle across dense, sparse, and condensed tiers. Crucially, we propose a Logits Calibration mechanism that normalizes diverse importance metrics into a unified probability space, enabling consistent Top-$p$ budgeting across heterogeneous heads. To bridge the efficiency gap, we offer a system-level solution, which rewrites fragmented, dynamic indices into contiguous physical layouts compatible with high-performance inference engine. Extensive experiments on math-reasoning benchmarks (AIME, U-Math) verify that HARD-KV achieves up to 2$\times$ throughput improvement over static baselines while maintaining high-fidelity generation in 10k+ token scenarios. Code is available at https://github.com/SuDIS-ZJU/HARDInfer.

We study the spectral perturbation of the empirical Fisher Information Matrix (FIM) of a parametric statistical model under two structured perturbations: departure of the input from a reference (in-distribution) ensemble, and finite-precision (quantized) perturbation of the model's parameters. For the first, under an explicit local curvature-monotonicity hypothesis on the dominant eigenvalue lambda_max of the FIM, we show departure from a reference manifold provably elevates lambda_max relative to a calibration baseline (Proposition 3.2), and discuss why this hypothesis is required, since curvature need not increase monotonically under every perturbation. Our principal result is a directional eigenvalue perturbation bound, via Weyl's inequality, showing lambda_max under a quantization noise perturbation is lower bounded by its unperturbed value up to a third-order remainder, and, under a mild genericity condition, strictly exceeds it at leading order (Theorem 4.3). We give two tractable approximations to lambda_max — one heuristic, one with a rigorous two-sided bound — and a completeness result for a threshold-based partition of an augmented state space. These results motivate using sigma_t = lambda_max(F_t)/lambda_base as a runtime monitoring statistic for deployed language models: the quantization result offers a mechanism for an empirical observation of our own, where a calibration threshold for this statistic was approximately 244 times larger than a preliminary full-precision estimate on a 4-bit quantized model, a single measurement rather than a value derived in closed form. We report supporting measurements (twelve models, n=1,080 trajectories) broadly consistent with our predictions, discuss the scope and limitations of every result, and state as an open problem the closed-form prediction of the quantization inflation magnitude our bound does not supply.