生成模型与LLM推理优化

When a camera revisits a previously observed region, a video world model should reproduce what was there before. This requires both remembering past observations and retrieving the right one for the current viewpoint. Key-value caches preserve visual detail but grow with video length; recurrent memory is compact but compresses history into a fixed-size state, so individual past observations are no longer directly accessible. We introduce LOCI, a hybrid spatial-memory architecture that keeps both representations. In half of the transformer blocks, main attention keeps a key-value cache of past observations; in the other half, it is restricted to the current chunk and complemented by a recurrent linear-attention memory whose reads and writes are conditioned on projective camera geometry, so viewpoint enters both memory addressing and stored content. Recurrent readouts flow into subsequent cache-backed blocks and supply their queries with accumulated scene context. On the public MIND memory benchmark and on held-out recorded trajectories, LOCI reproduces revisited content more faithfully than representative world models and a same-recipe full-softmax model; with full history, it lowers peak memory at equal length by about 30% relative to full softmax. With a bounded bank of retained observations, it streams long videos at constant memory and remains more faithful than full softmax under the same budget.

Fast matrix multiplication saves multiplications through exact cancellation, but rounding sums that mix token rows can leave contributions from later tokens in earlier language model outputs. This threatens prefix invariance, which multiple-choice likelihood scoring relies on: a scored likelihood must depend only on its allowed prefix. On Qwen2.5-14B-Instruct, two fast FP8 realizations repaired to ordinary-looking accuracy still change the answers chosen by likelihood on 5.83% and 10.00% of 240 OpenBookQA items when only the text after the allowed prefix is replaced with the bf16 model's own greedy continuation. Both row-local controls, the bf16 model and a deployed FP8 matrix multiplication kernel, change none. Accuracy thus does not certify prefix invariance, and the stability criteria we analyze cannot tell realizations apart: across all 512 sign variants of two-level Strassen they stay constant while teacher-forced perplexities span a 772.4$\times$ range on the same model. We therefore construct certified realizations of two-level Strassen on bounded integer codes that quantize token rows independently, then mix and cancel exactly before rescaling, using 49 block multiplications instead of 64. Our certificate guarantees bitwise equality to a prescribed row-local classical int8 operator at the same quantization specification, so every certified realization inherits its prefix invariance. Certification thus turns realization choice into a pure cost decision: which certified realization runs can no longer change a single scored likelihood.

Low-precision training rounds tensors that the backward pass reads again, often for several gradients; each use can read the forward's rounded value, the original, or a new random rounding. This backward-state policy looks like a memory and precision detail, settled by copy accuracy and final loss. We argue that it is part of the learning algorithm, and that neither check shows whether it is right. Copy accuracy does not decide the outcome: in three pairs of 390M runs with an emulated FP8 backward, training fails when attention's backward reuses the forward's rounded output and succeeds with a new rounding from the same distribution. Even the most accurate copy, the original itself, can be wrong by our reference: the gradient of the forward pass as it actually ran, with gradients passed through rounding unchanged. For example, a normalization output stored in low precision feeds two gradients: the gain's gradient needs the original, but the next layer's weight gradient needs the rounded value that layer multiplied. Final loss, the other check, does not rule out the error of reading the original for both: it persists in models trained with such a store, while planned loss comparisons stay within a margin fixed in advance. We therefore derive from this reference which value each use must read, or which substitute gives the same gradient on average with the forward held fixed, and check these per-use requirements on single operators, without training. In three tests using PyTorch and Transformer Engine, the requirements predicted beforehand whether reuse changes what the backward computes on average relative to an independent copy, and every prediction held. Backward-state policy is thus part of the learning algorithm: it should be specified and checked use by use, not settled by copy accuracy and final loss.

Scaling attention parameters can improve language model quality, but retaining full token histories makes additional heads costly at long contexts. Furthermore, since attention retrieves and combines contextual information, parameter scaling should also support longer contexts. We therefore ask whether attention parameter scaling can directly enable efficient and effective context scaling. We introduce NAMOH, an architecture-native sparse attention mechanism that activates $K$ of $H$ heads per token. Each head retains only its assigned tokens and performs causal attention within this subsequence. Head selection thus jointly determines active parameters and available context without scanning the full history. Under balanced assignments, increasing $H$ at fixed $K$ shortens head histories and reduces per-token key-value (KV) access without increasing total KV storage. We further support head-relative rotary position embeddings to shorten positional spans within routed subsequences, aiming to mitigate position-induced attention noise. Experiments show that NAMOH can outperform fully activated models with the same total parameters, while enabling more efficient long-context inference than smaller dense models with matched active parameter counts. It remains compatible with GQA and existing sparse attention mechanisms. We hope this work offers a new path for scaling attention, with parameter scaling directly enabling context scaling.

World-action models (WAMs) leverage pretrained video models to improve generalization in robot control by jointly predicting future visual states and actions. This capability comes at a substantial inference cost, as dense future-frame tokens are repeatedly processed during denoising. Prior methods address this by token pruning that prioritizes visual fidelity to reduce denoising costs in video diffusion models. However, these methods do not use action relevance to determine which future-frame tokens to retain during joint denoising in WAMs. In this paper, we propose Sparse-WAM, a training-free framework for action-guided sparse imagination that selectively processes future-frame tokens to accelerate WAM inference. We observe substantial overlap in the spatial distribution of attention from action tokens to future-frame tokens (action-to-future attention) between consecutive denoising steps, despite continued updates to the future representations. Motivated by this, we develop Action-Guided Token Selection to retain frame-specific action-relevant regions together with cross-frame context. However, a naive implementation can incur attention-scoring and token-packing overhead that offsets the computational savings from pruning. We therefore introduce Pilot, an efficient engine that reduces sparse inference overhead through lightweight scoring and cross-step reuse of token selections. On LIBERO with FastWAM-Joint and RoboLab-120 with Cosmos 3 Edge, Sparse-WAM achieves inference speedups of approximately $2.0\times$ and $1.8\times$, respectively, over dense eager inference on an NVIDIA RTX 4090, while largely preserving task performance.

Multimodal mixture-of-experts (MoE) models combine sparse expert activation with visual-language capabilities, yet their inference remains costly because long visual-token sequences repeatedly incur attention, routing, dispatch, and expert-MLP computation. Existing methods typically compress either the token or expert dimension, leaving redundancy along the other. Our analysis reveals two complementary regularities: the depth required for visual propagation varies across inputs, while text-token routing exhibits concentrated and recurrent expert-importance patterns. Based on these observations, we propose DecoMoE, a two-dimensional structured compression framework that decouples visual propagation from expert computation. The Sample-Adaptive Visual Boundary (SAVB) predicts an input-dependent visual-exit layer at which the visual-token block is removed. The Routing-Calibrated Expert Prefix (RCEP) reorders experts offline using text-token routed mass and, from this predicted exit layer onward, retains at each MoE layer the shortest contiguous prefix covering a target routed-mass fraction. We evaluate DecoMoE on Qwen3-VL-MoE and InternVL3.5-30B-A3B across six benchmarks. On Qwen3-VL-MoE, DecoMoE retains 97.91% of dense-baseline performance while reducing computation from 27.06 to 16.73 TFLOPs and latency from 0.44 to 0.26 seconds, yielding a 1.69x speedup. Code will be available at https://github.com/ShawnTan86/DecoMoE.

Diffusion Transformers (DiTs) have become a dominant architecture for video generation, but their efficiency is limited by the quadratic complexity of full attention. Sparse attention reduces this cost by retrieving important blocks and computing attention only within them, but inaccurate retrieval can either degrade generation quality or yield unnecessary computation. We identify two retrieval mismatches in methods that retrieve blocks using the averaged representations of query and key blocks: (i) query-side aggregation mismatch, where averaging queries before Softmax fails to preserve their individual attention preferences, and (ii) key-side clustering metric mismatch, where standard Euclidean clustering in the original key space can group keys with dissimilar QK scores under the current query, so their average representation may not accurately represent how the current query scores individual keys. These mismatches can lead to inaccurate block retrieval. To address these mismatches, we propose PARK, a training-free sparse attention method for accurate block retrieval. PARK retains every original query, independently normalizes its attention over key blocks, and then averages these distributions within each query block. It also uses information from the current queries to transform keys before clustering, so that keys receiving similar QK scores are grouped together. A fused GPU kernel further reduces the overhead of block retrieval. Experiments on HunyuanVideo and Wan demonstrate that PARK improves block retrieval accuracy and preserves generation quality while accelerating inference, achieving the best quality-efficiency trade-off among the compared sparse attention methods.

Test-time scaling has recently emerged as a powerful approach for improving LLM reasoning by allocating additional computation during inference, substantially enhancing accuracy on challenging tasks such as mathematics and coding. To accelerate the exploration of reasoning paths, recent studies proposed speculative execution. However, we show that supporting speculative execution poses two unique challenges for LLM serving systems: (1) an explosion in the search space of candidate paths and (2) frequent, fine-grained verification tasks for candidates. To address these challenges, this paper proposes SpecScale, a serving system for efficient speculative execution. We introduce three techniques to reconcile the trade-off between latency and computational overhead: (1) early pruning of low-quality candidate paths, (2) deduplicating computation across redundant candidate paths, and (3) deferring fine-grained verification tasks. We evaluate SpecScale on challenging reasoning benchmarks, including MATH and Olympiad. Our results show that SpecScale significantly outperforms both non-speculative and recent speculative approaches, delivering substantial improvements in throughput and latency while preserving answer quality.

Vision-language models face a fundamental scaling bottleneck: the number of visual tokens grows with both temporal duration and spatial resolution, making long-video understanding expensive for the vision encoder and the language model. Existing methods often compress visual tokens after dense encoding, creating a mismatch between the representation used during training and the compact interface required at deployment. We present CoVisco, a codec-native vision encoder with native token compression for unified image-video understanding. By combining codec-native input support with segmented attention, CoVisco can encode long visual inputs in a single forward pass without forming dense patch-to-patch interactions across all frames. Each temporal segment is equipped with learnable abstract tokens that learn a compact segment-level representation, while fine-grained patch tokens remain available throughout the encoder. Alternating intra-segment and abstract-communication layers preserve video-level context through the abstract-token channel. A lightweight selector further exposes either abstract tokens alone or abstract tokens augmented with a runtime-selected subset of patch tokens, yielding a compact visual interface that reduces the visual context and prefill burden of downstream MLLMs while retaining fine-grained evidence when needed. Pretrained with contrastive objectives on 565M image--text pairs and 6.4M videos, CoVisco shows competitive performance on video-oriented embedding and multimodal understanding benchmarks. In the evaluated four-segment, 64-frame setting, abstract-only inference uses only 400 visual tokens while achieving video-understanding performance close to, and on some benchmarks exceeding, OneVision-Encoder. Selected patch tokens further improve fine-grained video reasoning. Project URL: https://github.com/ernie-research/CoVisco.git

A transformer language model performs a bounded amount of computation per token, and recent work by Vishal Sikka, former CEO of Infosys, argues that this bound limits which tasks a model can carry out or verify (arXiv:2507.07505). We ask how much of the budget beneath that ceiling is spent on work the model has already done. Serving is stateless across requests: a model that answers a second question about a document recomputes the document's attention state from the first token. On seven real-world datasets, 98.7% of prompt tokens were text the model had already read. We present Galahad, a memory layer for vLLM, SGLang and llama.cpp that makes this reading a one-time cost. Taliesin saves the model's key-value (KV) state for a block of text and loads it on the next request that contains the same bytes, instead of recomputing it. Blaise keeps the documents themselves and passes the model only the section a question needs. On a recall test with 100 facts hidden in a 97,000-token corpus (Gemma 4 31B), Taliesin alone let the model attend to the whole corpus and answered 98 of 100 on llama.cpp at 3.0 s and 572 J per question, against 10 of 100, 9.3 s and 2,754 J for the same model without Galahad, which could hold only the last 12,000 tokens. With Blaise added, the model read about 668 tokens per question and answered 100 of 100 on all three runtimes at 0.59-0.64 s and 200-213 J; a tuned RAGFlow pipeline answered 77. Storing the corpus is a one-time cost of about 100 s and 28 kJ, whose energy is recovered after 13 questions. Restored state is bit-identical: all 262,144 output logits matched after restart, rehydration and hot-load. Galahad worked with all 30 models we tested under vLLM, and it fails closed: any load that does not pass its checks is recomputed. Together these results move LLM serving from stateless to stateful inference.

World-action models (WAMs) jointly generate future world states and actions through iterative denoising, using shared weights to process heterogeneous semantic streams of video, proprioceptive, and action tokens. Quantization reduces inference cost, but comparable numerical errors in different streams can have markedly different effects on final actions, making numerical accuracy alone insufficient for reliable control. We introduce SteerQuant, a 4-bit quantization framework for WAMs that steers errors toward computations with less influence on final actions. It maps how each stream's quantization errors affect final actions and uses this map to guide shared channel scaling. Activation scaling is further calibrated for each stream and denoising step to accommodate changes in activation ranges and action impact. This adapts quantization to different stream requirements without duplicating weights or increasing bit-widths for selected streams. To reduce the extra kernel launches and memory traffic introduced by scaling, we develop Rudder, a 4-bit inference engine for WAMs that fuses scaling and output compensation into low-bit kernels. Under W4A8 and W4A4, SteerQuant maintains mean LIBERO success within 0.8 percentage points of full precision, while delivering up to $2.23\times$ denoising speedup over BF16 across three WAMs with reduced peak GPU memory usage. On a real dual-arm robot, W4A8 deployment achieves a $1.35\times$ end-to-end inference speedup while maintaining average task success relative to BF16.

Large language model (LLM) inference is increasingly moving toward lower precision to realize the throughput of hardware accelerators, but aggressive post-training quantization (PTQ) can degrade model quality. We present QATFactory, an open-source framework for deployment-aligned quantization-aware distillation (QAD) and reinforcement learning (QARL). QATFactory simulates deployment-time quantization while performing matrix multiplications in BF16, allowing models to adapt to quantization noise without requiring training hardware that natively supports the target format; for example, it supports NVFP4 training on H100 GPUs, which lack FP4 Tensor Cores. The framework supports NVFP4, MXFP4, and $\text{llama}.\text{cpp}$'s Q4_K format; dense and mixture-of-experts models; and both full-parameter and LoRA-based training. It exports checkpoints directly to vLLM and $\text{llama}.\text{cpp}$ without an additional lossy conversion step or added inference overhead. With QATFactory, we conduct extensive experiments on models ranging from 8B to 230B parameters and evaluate exported checkpoints in production inference engines. Across models and formats, QAD consistently improves deployed-model quality over strong PTQ baselines. On Qwen3.5-9B, QAD achieves average benchmark accuracies of 68.9% under NVFP4 and 66.0% under MXFP4, outperforming the best PTQ results of 65.4% and 56.4%, respectively. Through our experiments, we found that although both FP4 formats quantize weights and activations at deployment, the best training strategy is format-dependent: NVFP4 generally performs better when only weights are quantized during training, whereas MXFP4 benefits from quantizing both weights and activations. At a fixed training token budget, training on fewer 32K sequences improves average accuracy by 1.9 points over training on more 4K sequences.

Modern transformers pair impressive capabilities with substantial memory and compute demands. Low-rank weight factorization reduces both while keeping the matrices dense, and thus efficient on standard hardware. Existing methods, however, choose the subspace to remove from each weight matrix with local closed-form criteria: activation energy, layer-wise reconstruction error, or a quadratic approximation of the loss. These criteria ignore how errors propagate through the network, so at high compression the errors compound with depth and performance collapses. We introduce Learnable Subspace Projections (LSP), which instead learns the subspaces to discard end-to-end. Each linear layer, or tied group of layers that read the same activations, is assigned an orthogonal projector. All projectors are optimized jointly against a global objective--the KL divergence to the dense model's output distribution or the model's original training loss--while the pretrained weights remain frozen. Projectors are initialized from a whitened SVD truncation, and ranks are allocated by the output KL each projector induces per parameter saved. After training, the projectors merge into standard low-rank factors, with each tied group sharing one factor. In attention, this also lets the model cache one narrow latent in place of full keys and values. Across LLMs (OPT-125M/1.3B, Qwen3-4B, Llama-2-7B) and ViT-B/16, LSP outperforms baselines, and its advantage widens as compression increases. At -70% compression, LSP brings Llama-2-7B to 10.9 WikiText-2 perplexity and 42.2% mean zero-shot accuracy, versus 13.3 and 36.0% for the strongest baseline. The factorized model decodes up to 1.6x faster than the dense model at small batch sizes, and aching the shared latent shrinks the combined memory of weights and KV cache by 13.5x at a 128k-token context, versus at most 6.5x for untied baseline factorizations.

When inference demand exceeds available compute capacity, model providers must decide which requests should be served first. Users have different tolerances for delay from an LLM API, but current priority pricing schemes compress these differences into coarse fixed-price service tiers. We design an inference auction that allows users to bid for faster service. Our auction allocates priority in an economically efficient way without sacrificing latency, and we develop fast algorithms for implementing prices that incentivize truthful bidding. We also design an autobidding agent for our inference auction, where users specify an inference budget and the autobidder dynamically adjusts its bids over time to maximize user utility subject to the budget constraint. Experiments validate the practicality of our auction: it increases system welfare while maintaining the cache utilization and latency advantages of SGLang, a state-of-the-art inference serving framework.

Long-context inference with Large Language Models (LLMs) is bottlenecked by the linearly growing memory of the key-value (KV) cache. Existing compression methods reduce the cache through token eviction or approximation, but degrade sharply at aggressive compression budgets. We propose PatchKV, a training-free framework that compensates KV cache compression methods by carrying part of the context in the model's weights. PatchKV pairs an off-the-shelf compressed KV cache with a context-specific weight patch, which is computed once at context-loading time and served for downstream queries for the context. The weight patch is derived in closed form via ridge regression, by aligning the block-wise activations of context-derived reference query tokens under the full cache and the compressed cache. Once merged into the model, the patch leaves the forward graph and per-query inference cost unchanged in the single-context, multi-query setting. Across long-context QA (SCBench with up to 170K tokens, SQuAD, NIAH) and math (GSM8K) benchmarks on three model architectures, PatchKV consistently improves cache compression methods, suggesting an alternative direction to compensate them at aggressive budgets.

As LLM capabilities advance, agents are tackling increasingly complex tasks over longer horizons. Their growing interaction histories make memory compaction essential for staying within context windows and reducing prefill cost. Existing methods summarize the history or compress its KV cache, often adding model computation to preserve information for future requests. A new user request can change which history matters, but reassessing that history with the model requires re-encoding it if the KV cache has expired. Past attention provides signals of historical importance and dependencies between messages, while relevance to the current task must be assessed using the new user request. We introduce ReCAP, a memory compaction method that stores attention-derived importance scores and dependency links in a lightweight, persistent context graph. For each new request, ReCAP combines stored importance with relevance cues from the request and follows dependency links to select messages and their supporting context, without additional model calls for selection. Compared with Codex's default summarization-based compaction, ReCAP reduces estimated latency for compaction and cold restoration by approximately 95% on both Qwen3-Coder and gpt-oss. It also roughly halves the historical context per call on SWE-Together at comparable task quality and improves accuracy on the code tasks of Lost-in-Conversation over full history by 19.8 and 41.2 points.

Autoregressive video diffusion supports streaming generation and interactive control, but its KV cache grows continuously with the generated history. Existing compression strategies either discard history using fixed windows or select tokens through local attention and similarity signals, which do not directly measure whether the current chunk contributes information beyond the retained context. We introduce DeCoPrune, a training-free method that treats cache compression as a denoising-consistency problem. We find empirically that denoising difficulty provides a useful proxy for a token's value in long-term retention: tokens with larger step-to-final discrepancies tend to carry visual evidence that is less predictable from the retained context. DeCoPrune measures each current-chunk token's denoising difficulty using the discrepancy between its intermediate clean prediction and final denoised value, retaining high-discrepancy tokens in the long-term cache while pruning those with low discrepancy. To evaluate information retention, we introduce CMBench, comprising 58 approximately one-minute generated or real-world context episodes and 116 Reappear or Revisit continuation tasks that require recalling specific previously observed objects or scenes. Experiments with LingBot World v2 show that DeCoPrune preserves near-FullKV long-range recall while pruning over 85% of historical KV tokens and accelerating continuation generation by over $4\times$, substantially outperforming the evaluated compression baselines at comparable budgets. These results indicate that denoising consistency can serve as a model-intrinsic signal for retaining long-range information while reducing autoregressive inference cost. Our project homepage is https://decoprune.github.io. The code is available at https://github.com/DeCoPrune/CMBench, and the benchmark at https://huggingface.co/datasets/Aoraku/CMBench.

Large language model (LLM) serving requires substantial memory to store model weights and KV caches. As models grow larger and contexts become longer, memory capacity and bandwidth increasingly become bottlenecks for serving performance. Agentic workloads compound this pressure through repeated interactions over growing contexts, making it increasingly important to retain KV state for reuse. High-bandwidth flash (HBF) offers a way to expand accelerator memory capacity for large language model (LLM) serving, but its access costs and limited write endurance complicate its use. We evaluate HBF for high-throughput agentic serving across system design and scheduling choices to understand when additional capacity improves serving performance and energy efficiency. We introduce an HBM-HBF-host hierarchical storage system and buffered cache-aware scheduling, and use trace-driven simulations to analyze their effects on performance, energy consumption, and HBF write lifetime. Across the evaluated workloads, the fastest HBF-augmented systems reduce completion time by 36.1-87.7% relative to HBM-only systems. Modeled energy savings reach 59.1%, with benefits depending on the workload and weight placement. Buffered cache-aware scheduling extends estimated HBF write lifetime from 1.21 to 14.82 years in the evaluated configuration. These results demonstrate the importance of coordinating data placement and scheduling to improve serving efficiency while sustaining a practical HBF write lifetime.

We present Cascadia, a system for serving large language models on fleets of commodity Intel AIPCs using their CPU, integrated-GPU, and NPU resources. Every node embeds ingress, scheduling, and execution; inference requests require no dedicated routing control plane. Nodes join a libp2p QUIC mesh using CA-issued ed25519 admission certificates, gossip signed capabilities, exchange live load over direct peer streams, and route OpenAI-compatible requests to eligible peers. An operator-run certificate authority handles admission and fleet management outside the inference path. Three serving modes share one interface: whole-model execution on one node, load-balanced replicas, and pipeline-sharded chains using the compilation and speculative decoding mechanism of our companion paper. Optional KV-cache mobility reuses compatible conversation prefixes after a routing move, with cold recomputation on a miss. Signed response receipts and hash-chained logs support provenance and audit. A three-node Phi-3.5-mini NPU testbed delivered 3.10x the response throughput of its one-node configuration under ten concurrent requests; a separate four-node deployment recorded 4.06x the throughput of direct single-node serving. Paired latency observations, runtime measurements, and internal functional checks characterize the tested configurations. We compare Cascadia with IBM, Nutanix, VMware, and HPE platforms on deployment footprint, hardware requirements, scheduling, scaling, licensing, and trust, using vendor documentation. The paper repository provides benchmark scripts, curated measurements, and a claim-to-evidence map.

LLM inference systems may vary batch composition, prompt chunking, prefill/decode execution, and KV-cache reuse, eviction, or recomputation. These optimizations should not affect system outputs. Production systems, including vLLM's batch-invariant mode and SGLang's deterministic mode, target this goal but lack a formal system-level specification. We formalize deterministic LLM inference: under a fixed model and deployment configuration, requests with the same prompt and initial sampler state produce bitwise-identical logits at corresponding output positions across executions. Our tests find that these production modes produce different logits under some execution variations. To address this limitation, we present Vosti, an inference engine designed and verified against this specification. Vosti chooses kernels independently of runtime engine state and ties cached KV values to their logical token prefixes. Its proof decomposes at the engine/GPU kernel boundary: a Verus inductive proof establishes that scheduling and the paged, prefix-sharing KV-cache preserve output logits, while a Triton analyzer proves bitwise-equal selected kernel outputs across batches, query lengths, and paged KV-cache layouts. Vosti produces bitwise-identical logits across every tested execution variation and achieves performance comparable to vLLM's batch-invariant mode on decode-heavy workloads, while providing a stronger, formally verified determinism guarantee.

Production LLM serving stacks combine an inference engine's local prefix cache with a shared KV-cache tier for fleet-wide reuse. The local cache distinguishes requests by adapter, weight configuration and sharing domain, but the shared tier may key entries only by token content and coarse model metadata. This boundary erases provenance and lets identical tokens under incompatible computational or sharing contexts collide. We call this composition gap provenance-blind reuse and present its first systematic study. A source audit of three vLLM connectors confirms the structural omission, while runtime experiments reproduce it across vLLM and two SGLang releases, 12 models from 7 families (0.5 B-32 B), and over 160 configurations. Cross-adapter collisions reduce accuracy from 0.94 to 0.64, incompatible KV representations reduce reasoning accuracy to zero, and salt omission enables 93% prompt identification from timing. We formalize the missing guarantee as the KV provenance contract: for a declared dimension registry, shared keys must be injective over computational and sharing provenance, with identities stable across workers. Any dimension with a stable identity can therefore be added without connector-specific key logic. A canonical descriptor binds per-request and per-worker provenance into lookup and store keys, while a differential checker detects dimensions that change KV state without changing the key. Implemented in vLLM and SGLang 0.5.20 across three cache paths, provenance binding eliminates unsafe reuse while preserving legitimate sharing. Hit-path latency changes remain within 0.34 ms and below run-to-run variation; retention grows with provenance diversity.

Sparse attention is widely used to accelerate long-context inference in modern large language models (LLMs), but its input-dependent execution behavior introduces previously unexplored privacy risks. We identify a new GPU micro-architectural side channel, termed Sparsity-Induced Memory Access (SIMA), which arises from secret-dependent key-value cache access patterns induced by sparse attention. Based on this observation, we present SparLeak, a phase-aware side-channel attack that extracts SIMA traces during LLM inference and enables two practical privacy extractions: query attribute inference from prefill-phase traces and autoregressive response reconstruction from decoding-phase traces. By reconstructing approximate token-level sparsity profiles from page-level observations and applying profiling-based learning, SparLeak accurately recovers sensitive information, including user-query attributes and private LLM response content. Extensive evaluation across three LLM architectures, three sparse attention mechanisms, and three privacy-sensitive datasets shows that SparLeak achieves average attack success rates of 90.9% for attribute inference and 87.3% for response reconstruction under real-world LLM serving settings, highlighting the significance to account for SIMA leakage when deploying sparse-attention-based LLM systems. We provide anonymized SIMA traces, trained attack models, evaluation scripts, and documentation as artifacts at https://anonymous.4open.science/r/Janus_artifacts/.

Long-context LLM agents accumulate interaction histories that strain KV-cache memory and attention computation. Although sparse attention reduces these costs, heterogeneous cache representations and workflows hinder integration with existing inference engines, while prior sparse-serving abstractions support only specific layouts or workflows. We present SparseEngine, a ground-up, sparse-first inference engine whose shared lifecycle contract lets each method control its KV representation and computation while coordinating state transitions with common serving infrastructure. SparseEngine supports 15 methods across four categories and enables cross-request state management through Chain Cache, which resumes KV-eviction methods from retained history, and controllable Prefix-Cache Pruning, which removes KV from selected history regions while preserving logical-prefix matching. While maintaining method quality, SparseEngine delivers over 10x higher throughput with KV eviction, over 2.5x faster decoding at matched concurrency than vLLM, and over 2x end-to-end speedup on agent benchmarks. The code is available at https://github.com/CURRENTF/SparseEngine.

Post-training quantization (PTQ) has become a widely adopted technique for reducing the memory footprint and inference cost of large language models (LLMs). However, recent studies reveal that when applied to reasoning models, PTQ not only degrades reasoning performance but also exacerbates overthinking, leading to longer reasoning trajectories. These issues may offset the efficiency gains expected from lower-precision inference. Existing approaches mainly rely on complex optimization procedures. More recent lightweight inference strategies instead use predefined overthinking markers, limiting their adaptability across quantized models. To address these issues, we propose Reasoning Analysis and Token-level Inference Optimization (RATIO), a framework that identifies model-specific overthinking tokens and assigns each a tailored penalty. RATIO first introduces Quantization-aware Reasoning Behavior Analysis (QRBA) to identify overthinking tokens by analyzing discrepancies between full-precision and quantized models. It then adopts Token-Specific Penalty Determination (TSPD), which leverages full-precision guidance to derive token-specific penalties without additional training. Extensive experiments show that RATIO achieves a better accuracy-efficiency trade-off than existing token-level interventions. Specifically, RATIO achieves up to 9.8 points accuracy improvement and reduces chain-of-thought (CoT) length by up to 51.3% compared with quantized baselines. The code will be available at https://github.com/steven-bao1/RATIO.

Mixture-of-Experts (MoE) architectures allow frontier language models to scale to trillions of parameters, but their deployment is constrained by massive memory footprints and memory-bandwidth limitations. Although modern accelerators provide Sparse Tensor Cores (SpTCs) that reduce weight storage and increase throughput through low-precision semi-structured sparsity, exploiting them for MoEs remains challenging because of substantial model-quality degradation and the lack of grouped sparse GEMM primitives. We present an end-to-end hardware-software co-design framework that compresses expert weights into hardware-native, low-precision sparse representations and accelerates their execution on SpTCs. Algorithmically, our framework relaxes discrete semi-structured support selection through continuous reparameterization, enabling differentiable joint optimization with quantized weights under a router-weighted reconstruction objective and scalable expert-parallel compression. Systemically, we develop a custom grouped sparse GEMM kernel tailored to low-precision sparse MoE inference on SpTCs. Across MoE models ranging from 30 billion to one trillion parameters, our framework improves state-of-the-art joint sparse-quantization accuracy by up to 4.35 percentage points while preserving 96.09% of the original model's performance. On NVIDIA B200 GPUs, our kernel outperforms the vendor baseline by up to $1.65\times$, increasing serving throughput by $1.18\times$ and reducing end-to-end latency by up to $4.03\times$. These results establish hardware-software co-design as a practical path toward scalable and efficient MoE deployment.

Deploying large vision-language models (VLMs) onboard satellites enables onboard data processing and reduces raw data downlink. However, onboard inference faces two resource challenges. Limited onboard memory and energy require model compression and distributed deployment. Dynamic resource availability requires fast deployment decisions as illumination, battery levels, and communication conditions change. We present SCORAS-MoE, a joint compression and deployment framework for mixture-of-experts (MoE) VLMs in low Earth orbit (LEO) satellite networks. To address limited resources, SCORAS-MoE measures the perturbation of the routed MoE output caused by low-rank approximation, assigns higher ranks to more sensitive experts, and distributes compressed model shards across satellites for cooperative inference. The compressed models yield profiles of measured accuracy and inference energy. To adapt to dynamic resources, the online scheduler selects profile compositions and shard placements in each slot. For each candidate composition, it reduces placement to a minimum-cost assignment problem solved by the Hungarian algorithm, while enumerating the compositions yields the optimal deployment for the current-slot objective. Experiments on Qwen3-VL-30B-A3B-Instruct show that allocating ranks based on output perturbation is particularly effective under aggressive compression, with an absolute gain of $3.7%$ in mean accuracy over uniform rank allocation when expert projections retain $30%$ of their original parameters. The fixed-profile scheduler achieves higher throughput with fewer service switches and lower battery impact than the evaluated proximal policy optimization (PPO) and evolutionary baselines, with respective speedups of $8.7\times$ and $183.5\times$. Adaptive profile selection further improves the balance between service quality and energy use.

System One models such as Jev offer an efficient alternative to generative language models for tasks that require decisions rather than open-ended responses. However, existing Jev models exhibit limited Chinese-language decision accuracy, restricting their utility in both general and specialized settings. In this paper, we introduce Chinese-Jev, a System One model that addresses this gap through a unified data processing and training pipeline. Our data processing protocol converts heterogeneous Chinese-language annotations into probability targets over candidate options, enabling a shared training formulation across domains and question formats. To enable efficient inference, Chinese-Jev adopts a lightweight encoder-only backbone for text encoding and learns to score candidate answers through decision-oriented training. To address the misalignment between the pre-training distribution and downstream Chinese-language scenarios, we first train the model on a general-purpose corpus of 10 million examples, then fine-tune it separately for the medical, legal, and financial domains. To evaluate decision accuracy and calibration in both general and domain-specific Chinese-language settings, we introduce Chinese-Jev Bench (CJ-Bench). After first-stage pre-training, Chinese-Jev exceeds the accuracy of the closed-source Jev model by 1.24% on general-domain tasks while achieving a 20.3x speedup. Subsequent domain-specific fine-tuning yields a 4.0% accuracy improvement over Jev in medicine and achieves 92% of Jev's average accuracy across specialized domains, with a 17x speedup and an average latency of only 15 ms per example. We further demonstrate on-device deployment of an INT8-quantized model on mobile devices, achieving an inference latency of approximately 1.0 second per decision. The project is available at https://gulucaptain.github.io/Chinese-Jev/.

On-policy distillation (OPD) trains a student model on its self-generated trajectories with dense token-level teacher feedback. However, naive OPD may suffer from teacher unanchoring, where the student's reasoning trajectory drifts far from the teacher, causing the teacher to be queried on states it would hardly visit and thus provide unreliable supervision. We propose Interactive-Policy Distillation (IPD), which applies adaptive teacher intervention to the student rollout. Under a bidirectional propose-and-verify state machine, the student and teacher alternately exchange their roles as proposer and verifier, and collaboratively generate mixed-source trajectories. Then different supervisions are applied according to the source of each token. This bidirectional propose-and-verify mechanism and the source-split loss make IPD not only a more performant distillation method, but also a unified bridge between on-policy and off-policy paradigms. To make the interleaved dual-model rollouts more efficient, we also design a dedicated fused inference engine that co-hosts both models in one serving instance with separate KV caches and instantiates the state machine model to distribute, collect, and process requests. On math reasoning tasks and across multiple teacher-student model pairs, student models trained with IPD not only outperform those trained with OPD, but also demonstrate higher data efficiency. Specifically, when distilling Qwen3-30B-A3B into Qwen3-1.7B-Base, IPD brings a +3.28 mean@8 and a +3.28 best@8 benchmark-averaged accuracy improvement compared with OPD. Besides, IPD only consumes about 1/4 of the training examples and steps to outperform OPD trained on the whole training dataset for one epoch. We also investigate the impact of different loss variants and takeover / handback configurations, and demonstrate the robustness of IPD on different training data.

End-to-end attention-based speech recognition is accurate offline but hard to stream: outputs can depend on future audio, and a little future context per layer makes the lookahead grow with the number of layers. We address this with two mechanisms. A bounded-lookahead chunk encoder caps every chunk's future receptive field at a constant number of chunks, independent of the number of layers, via one age-selection rule shared by self-attention and the depthwise convolution. On this encoder, dynamic future-chunk decoding lets a per-token trigger commit a token or wait and re-decode it; we propose a learned trigger as the general mechanism, with a simple confidence threshold as an effective fallback. On full LibriSpeech test-clean the dynamic system matches the best static-lookahead accuracy (6.5%) at a median latency of 306 ms versus 860 ms for one-chunk static lookahead, and a wait budget bounds the deferral tail below the static baseline's 90th percentile at 0.1 points more WER.

Softmax attention is ubiquitous in modern machine learning, but its quadratic scaling with sequence length makes it costly. To reduce this cost, attention is often approximated with fast algorithms, which incur error but can still perform well in practice and on some inputs. At the same time, the growing diversity of attention applications makes approximation guarantees that do not depend on particular input structure a compelling target. For such uniform guarantees over all inputs, known runtime lower bounds rule out fast algorithms for near-exact attention, but leave open the practically important regime: is there an efficient algorithm with even a modest uniform approximation guarantee? We answer this question negatively. Under standard complexity-theoretic assumptions, no truly subquadratic algorithm can approximate attention with any nontrivial additive or relative guarantee uniformly over all inputs. This impossibility holds in the mildest parameter regime for which known algorithms do not already achieve strong approximation guarantees in near-linear time, and extends to practically relevant relaxations: even after polynomial preprocessing of the KV cache, no efficient algorithm can obtain a nontrivial uniform approximation guarantee, or identify a small set of keys receiving substantial attention under sparsity. Overall, our results settle the computational limits of uniform attention approximation.