生成模型与LLM推理优化

Modern KV cache management assumes the chatbot workload: prompts arrive once and the cache grows append-only, so prefix caching and forward-only eviction are correct by construction. Agentic LLMs break this assumption. Their conversations evolve through policy-driven editing: failed tool calls are retried, stale outputs dropped, trajectories pivoted. Two distinct cache problems result. First, identical content moves to new positions between turns, invalidating exact-prefix caches even though the underlying KV would still be valid; recent work on position-independent caching for MLA addresses this reuse problem. Second, and this paper's focus, a policy may need to direct the serving system to actively remove or replace a span of cached content and continue without re-prefilling everything that came after. No existing primitive offers this. Production agentic harnesses fall back to re-prefill on every edit, paying full prefix-recomputation cost; kernel-level eviction methods make their own decisions and cannot accept policy directives from outside the kernel. We introduce Leyline, a serving-side primitive that closes this gap. A declarative directive 4-tuple separates what to edit from how to preserve position correctness. The policy declares the edit and its mode (in-place splice or prefix-trimmed re-prefill for semantic forgetting); an architecture-agnostic interface routes to a per-architecture kernel that restores attention math via a closed-form RoPE-rotation correction. The splice kernel lifts replay cache-hit by +11.2 pp and cuts latency by up to 241 ms. A ten-line truncation rule routed through the same interface lifts agentic solve rate by +14.3 pp on debug-gym. The mechanism is open; the policy space it enables is the agenda.

We present a post-training quantization (PTQ) approach for Wan2.1-T2V-14B, a 14-billion-parameter text-to-video diffusion transformer, targeting the W8A8 HiFloat8 (HiF8) format on Ascend 910B NPUs. A central challenge in quantizing video DiT models is the heterogeneous activation distribution across transformer blocks: boundary blocks (the first and last few blocks) exhibit fundamentally different statistical properties from middle blocks, making uniform quantization ineffective. We conduct a systematic per-block activation analysis across all 40 WanAttentionBlocks and use the findings to motivate a boundary-protection strategy that retains the first two and last three blocks in BF16 while quantizing the remaining 35 blocks with W8A8 HiF8. The proposed PTQ method matches or marginally exceeds the BF16 baseline on all five VBench dimensions evaluated, indicating no measurable accuracy loss within the 5-prompt evaluation set. An ablation study over four protection configurations confirms that full boundary protection yields the highest average VBench score, validating the data-driven block selection. We additionally investigate quantization-aware training (QAT) as a complementary fine-tuning stage and analyze the conditions under which it fails to outperform plain PTQ on single-card hardware.

Efficiently serving large language model (LLM) inference tasks is crucial both for user-perceived latency such as time-to-first-token (TTFT) and for GPU utilization. However, LLM request routing, that is, assigning each inference request to a GPU instance, is particularly challenging: execution is highly input-dependent; batching and KV-cache reuse create strong cross-request coupling; and latency responds nonlinearly to context length, model/engine settings, and heterogeneous accelerators. As a result, simple traditional load balancing algorithms, and even heuristics tailored for LLM inference, fail to achieve good performance. We present Lodestar, a novel learning-based request routing system for distributed GPU clusters. Lodestar continuously collects a snapshot of the cluster at per-request level, including real-time instance state, request characteristics, and observed performance, and trains an online reward predictor that it uses to route inference requests to the instance that will maximize given reward (e.g., minimizing TTFT). Lodestar is cloud-native and works seamlessly with existing serving stacks (vLLM). With continuous online adaptation to changing workloads and infrastructure conditions, Lodestar achieves 1.41x lower average TTFT and 1.47x lower P99 TTFT on average (up to 2.15x/1.86x on homogeneous and 4.38x/4.42x on heterogeneous clusters) compared to a state-of-the-art prefix cache and load-aware heuristic, and learns these efficient routing strategies within about 5 minutes, based on experiments in a public cloud GPU cluster.

Speculative decoding accelerates large language model inference by pairing a target model with a lightweight draft model whose proposed tokens are verified in parallel. A common way to build draft models, like EAGLE3 or DFlash is supervised fine-tuning (SFT) on target-generated trajectories. However, we observe that SFT quickly plateaus: the draft model's acceptance length on test data stops improving. The reason is an offline-to-inference mismatch: In SFT, the drafter learns from fixed target-generated trajectories, whereas during speculative decoding it is evaluated on blocks proposed under its own policy. This motivates on-policy distillation (OPD), where the target model supervises the drafter on draft-induced states. Yet OPD remains difficult for draft models, as they cannot reliably roll out complete sequences independently, whereas target-assisted generation makes the collected sequences follow the target distribution and thus eliminates the on-policy signal. We therefore propose Draft-OPD, which uses target-assisted rollout for stable continuations and replays drafting from the verification-exposed error positions. This allows the drafter to learn from target feedback on both accepted and rejected proposals, focusing training on the draft-induced errors that limit speculative acceptance. Experiments show that Draft-OPD achieves over $5\times$ lossless acceleration for thinking models across diverse tasks, improving over EAGLE-3 and DFlash by 23% and 13%.

Real-time inference of vision-language-action (VLA) models is essential for robotic control. While visual token pruning has shown strong potential for accelerating inference, most existing methods mainly base pruning decisions on shallow-layer cues and risk discarding visual information required by deep layers. To address this issue, we propose SAFE-Pruner, a plug-and-play pruning framework that incorporates attention cues of future layers into pruning decisions. Specifically, we identify semantic attention consistency, the tendency that VLA models concentrate their attention probability mass on the same semantic entity across execution steps. Based on this observation, we design a forward-looking strategy to forecast the token saliency in deep layers, which prevents the premature removal of critical tokens and leads to more stable acceleration. We further introduce an adaptive subtask division strategy to detect abrupt attention shifts, thereby improving forecasting accuracy and pruning reliability. Extensive experiments in simulation and real-world settings demonstrate that our method achieves up to 1.89x speedup with a minimal degradation in success rate of less than 1.7%, while outperforming state-of-the-art methods by up to 1.9%.

Vision-Language Models (VLMs) process thousands of visual tokens per image alongside comparatively few text tokens, yet existing compression methods treat both modalities uniformly. We observe that the two modalities have fundamentally different properties: vision tokens are spatially redundant and dominate prefill, while text tokens are causally dependent and accumulate during decoding. Based on this asymmetry, we propose and empirically evaluate AsymVLM, which applies aggressive pruning to vision tokens before prefill using a learned importance scorer with per-sample adaptive budgeting, and temporal threshold-based eviction to text tokens only when they exceed a fixed budget. Our experiments indicate that AsymVLM achieves the highest FLOPs savings (up to 54%) among state-of-the-art methods while outperforming existing approaches by 2--3% on document and chart understanding tasks where visual information is spatially localized and query-specific, and maintaining competitive accuracy on holistic benchmarks. In text-dominated scenarios, our eviction strategy substantially outperforms standard LLM cache compression methods by adapting to the short-context nature of VLM.

Block-diffusion drafters have recently emerged as a powerful alternative for speculative decoding by predicting multiple future-token distributions in a single parallel step. However, since these parallel predictions are sampled from position-wise marginals rather than fully conditioned sequences, committing to a single greedy path often fails to capture the target model's preferred trajectory. To address this, we propose BASTION, a budget-aware speculative decoding framework with tree-based diffusion drafting. Unlike existing methods that rely on static tree topologies, BASTION dynamically constructs query-dependent trees by balancing draft quality against hardware constraints. Our framework integrates three synergistic components: (1) an acceptance surrogate that estimates expected accepted length via path confidence, (2) an online latency estimator that calibrates a hardware-aware roofline model, and (3) an adaptive best-first expansion that grows the tree until marginal gains no longer justify incremental verification costs. BASTION is training-free, preserves the target model's distribution, and requires no per-setting tuning. Across diverse benchmarks and GPU architectures, BASTION achieves up to a 6.61x speedup over standard autoregressive decoding, outperforming state-of-the-art block-diffusion baselines by 39%.

Key-Value (KV) cache remains a major bottleneck for deploying Large Language Models (LLMs) in long-generation tasks. Prior work often applies uniform compression across both prefill and decoding caches, but compressing the prefill cache degrades performance by corrupting critical context. While preserving the prefill cache is essential, decoding-phase compression remains underexplored, with existing methods relying on rigid recency windows or instantaneous attention. Our analysis of attention dynamics reveals strong temporal patterns: critical tokens receive sustained attention over long horizons, while local reasoning involves short-lived bursts. Static heuristics fail to capture this behavior, leading to premature eviction of important tokens or retention of stale ones. We propose Moment-KV, a decoding-time KV cache compression method based on momentum-driven temporal attention aggregation. Our method models token importance as a continuously evolving state, where attention is aggregated with decay, capturing both long-term influence and recent relevance. Experiments show that Moment-KV significantly improves generation fidelity in long-generation tasks (2.3-3.2 %) while maintaining decoding latency.

As large language models continue to scale, low-bit weight-only post-training quantization (PTQ) offers a practical solution to their memory-efficient deployment. Although block-wise PTQ is capable of matching the full-precision (FP) baseline on basic language modeling and understanding, its quality is degraded for generative tasks — especially at longer responses and extended chains of thought, which is critical in boosting task accuracy. We attribute this shortfall to two factors: (i) the omission of the unembedding layer (the LM head) in block-wise optimization and (ii) the reliance on the mean squared error (MSE) objective. Both factors cause the token probability distribution of the quantized model to misalign with that of the FP model, yielding notable accuracy drops on text generation benchmarks. To rectify the discrepancy, we introduce Logit-aware Final-block Quantization (LFQ), a simple yet effective enhancement to block-wise PTQ that quantizes the final Transformer block by minimizing the cross-entropy between the logits of the FP model and those of its quantized counterpart. By aligning token probabilities at the logit level in the final block, LFQ consistently improves the accuracy of complex generation tasks over state-of-the-art block-wise PTQ across diverse model families, while maintaining parity with FP baselines on language modeling and understanding.

Diffusion language models (dLLMs) generate text by iteratively denoising multiple token positions in parallel, offering an attractive alternative to strictly autoregressive decoding. In practice, however, block-wise dLLM inference exposes a difficult granularity trade-off: small blocks preserve local conditioning but require many denoising steps, whereas large blocks expose more parallelism but can make premature commitments and accumulate cache error. Existing acceleration methods typically choose a single block size per request, leaving the complementarity among block sizes unused. We show that block size itself is a useful branching dimension. Different block sizes induce related but non-identical KV-cache trajectories: branches often share an initial prefix, bifurcate at semantically decisive positions, and later agree on syntactically lightweight tokens. Motivated by this structure, we propose BlockBatch, a training-free online inference framework that executes multiple block-size branches for the same request inside a batched forward pass. BlockBatch coordinates these branches through confidence-gated token merging, leader-based synchronization, and periodic full-sequence refreshes that re-anchor local block updates to a globally consistent KV state. Across 3 representative dLLMs and 4 datasets, BlockBatch reduces denoising NFEs by 26.6% on average and achieves a 1.33$\times$ average end-to-end speedup over Fast-dLLM while preserving accuracy. These results identify block-size diversity as a practical and previously underexplored axis for branch-parallel dLLM inference.

Autoregressive (AR) video generation has emerged as a promising paradigm for long-horizon video synthesis, where each frame is generated conditioned on previously generated tokens. To accelerate inference, the KV cache is used to avoid redundant recomputation across generation steps. Nevertheless, its growth with generation length introduces increasing memory and error accumulation, limiting the scalability of AR models to even longer sequences. Existing KV cache compression methods mitigate this issue by selectively retaining only video tokens deemed important. However, most existing methods assess token importance using short-horizon signals derived from the current or historical generation context, making these methods prone to overlooking tokens that appear unimportant at early steps but later become critical for future frames. In this work, we identify an important property of trained AR video models: although RoPE-modulated queries evolve across autoregressive steps, the underlying canonical pre-RoPE query distribution remains remarkably stable throughout the video generation process. This approximate stationarity implies that future query distributions are estimable from historical statistics, enabling principled future-aware cache decisions without any additional training. Building on this insight, we propose Future Forcing, a training-free future-aware KV cache policy for AR video generation. Specifically, Future Forcing first constructs a future query proxy from historical statistics, then scores KV cache tokens by their importance under this proxy, and finally merges redundant token pairs within the affine subspace induced by the future query. Extensive experiments show that Future Forcing improves long-horizon consistency under limited KV caches, achieving up to 1.49 improvement in subject consistency on VBench-Long for 60s generation over existing AR video KV cache policies.

Video large language models (Video-LLMs) have demonstrated strong capabilities in video understanding tasks. However, their practical deployment is still hindered by the inefficiency introduced by processing massive amounts of visual tokens. Although recent approaches achieve extremely low token retention ratios while maintaining accuracy comparable to full-token baselines, most of them perform compression only at the late stage of prefilling, leaving the efficiency of the vision encoder unoptimized. In this paper, we first show that vision encoding contributes a large portion to the time-to-first-token (TTFT). Therefore, instead of compressing visual tokens only after the vision encoder, performing compression inside the encoder still leaves substantial room for exploration. Based on this insight, we propose EarlyTom, a training-free token compression framework that performs early-stage visual token compression inside the vision encoder, enabling significantly better TTFT reduction and higher throughput. In addition, we introduce a decoupled spatial token selection strategy that improves the overall compression effectiveness. EarlyTom reduces TTFT by up to 2.65x and FLOPs by up to 61% on a single NVIDIA A100 GPU for the LLaVA-OneVision-7B model, while maintaining accuracy comparable to the full-token baseline. These improvements substantially enhance the practicality of deploying Video-LLMs in real-world production scenarios.

Long-rollout causal video diffusion has converged on a fixed-size sliding-window KV cache, with recent progress innovating within this layout by changing which tokens occupy the window or how their positions are encoded. The per-head KV layout itself, a dominant contributor to streaming memory and latency, has been mostly left unchanged. In this paper, we present the first study of Multi-Head Latent Attention (MLA) in video diffusion. VideoMLA replaces per-head keys and values with a shared low-rank content latent and a shared decoupled 3D-RoPE positional key, reducing per-token KV memory by 92.7% at every cached layer. We further investigate why MLA succeeds in video diffusion even though the spectral assumption often used to motivate it in language models does not hold: pretrained video attention is not low-rank, with 99%-energy effective rank far above any practical latent dimension. VideoMLA retains quality at compression ratios where direct spectral approximation would predict large reconstruction error. We show that the MLA bottleneck, rather than the pretrained spectrum, determines the effective rank: both spectral and random initialization occupy nearly the full rank budget from initialization, and training preserves this budget while adapting within it. On VBench, VideoMLA matches short-horizon streaming video diffusion baselines, achieves the best overall score at long horizons among evaluated methods, and improves throughput by 1.23x on a single B200.

Autoregressive video diffusion models generate streaming video by producing frames sequentially, conditioning each chunk on previously generated content. These models are structurally anchored to the first frame: its key-value representation occupies a privileged position in the attention cache and serves as the primary scene reference throughout generation. As the cleanest and most error-free position in the cache, this anchor draws disproportionate attention, suppressing video dynamics, and locking scene composition to the initial viewpoint even as the scene naturally evolves. The result is a temporally shallow video in which motion, camera movement, and scene progression are dampened in favor of static consistency. To address this, we replace the static anchor with an adaptive state, a hidden latent that the model denoises alongside content at every chunk but never renders. Rather than referencing a frozen first frame, the model generates its own scene anchor at each step by attending to both the previous state and the current content, producing a reference that evolves with the generated content. Unlike standard video generation, which encodes an absolute notion of time, our formulation treats time as relative: every generation step sees the same positional structure regardless of how far generation has progressed, and the state transition is identical at every chunk. Together, these properties introduce a recurrence into the generation process, where denoising serves as the transition function, and the KV cache serves as the carrier, requiring no external module. Experiments demonstrate that the adaptive state substantially improves video dynamics, enabling richer motion and natural scene progression within generated videos.

Scaling Diffusion Transformers to generate high-resolution, long videos is constrained by the quadratic cost of self-attention, and existing sparse attention methods degrade under high sparsity. We show empirically that generation quality is determined not by the sparsity ratio itself, but by how well the sparse mask aligns with the tile-wise geometry of full attention. Based on this insight, we propose Veda, a distilled sparse attention framework that formulates tile selection as an explicit reconstruction problem from full attention. Veda integrates statistics-aware tile scoring with head-aware tiling to reduce estimation error and structural mismatch, enabling aggressive sparsity. A hardware-efficient tile-skipping kernel converts theoretical sparsity into practical wall-clock speedups. Experiments on large video diffusion models, including Waver and Wan2.1, demonstrate substantial acceleration with no noticeable degradation in generation quality. To generate 720P 10-second videos on Waver-T2V-12B, Veda achieves a 5.1$\times$ end-to-end speedup and a 10.5$\times$ self-attention speedup, reducing attention overhead from 92% to 50%. Notably, the gains increase with sequence length, indicating that Veda scales favorably with spatiotemporal resolution across models.

Multimodal Large Reasoning Models introduce the reasoning paradigm, demonstrating strong capabilities on complex vision-language tasks. However, they still suffer from severe hallucinations. Existing training-based methods typically mitigate hallucinations through response-level direct preference optimization (DPO), where the Chain-of-Thought (CoT) and the final answer are treated as a monolithic output and optimized jointly. We reveal that this formulation performs similarly to answer-only optimization, suggesting that it primarily learns answer-level preference, while leaving CoT-level supervision insufficiently exploited. To address this issue, we explicitly formulate a CoT-oriented preference term and derive Reasoning-Conditioned Direct Preference Optimization (RC-DPO), which models the CoT as a condition for answer generation and contrasts the preference for the same preferred answer under different CoT conditions, promoting answer-supportive reasoning chain alignment. To further improve optimization, we introduce a reasoning-enhanced preference data generation strategy that employs Monte Carlo Tree Search to discover visually grounded and logically consistent CoTs as positive samples, and attention-guided CoT token pruning to construct negative ones. Extensive experiments across various models and benchmarks show that RC-DPO effectively mitigates hallucinations and improves the reliability of the multimodal reasoning process.

Data-aware post-training quantization (PTQ) minimizes a per-token reconstruction loss on a small calibration corpus, implicitly weighting positions by their empirical frequency. For Automatic Speech Recognition (ASR), this misaligns with tail-sensitive risk: names, numerals, and domain-specific words receive proportionally little calibration mass. We propose Tail-Aware Reconstruction Quantization (\TARQ), a label-free PTQ framework that shifts calibration toward the lexical tail via \rareBAL, a closed-form per-Linear-layer rule equalizing common/tail mass, paired with a metric-consistent residual correction. \TARQ\ requires no entity labels, no curated calibration set, no validation decoding, and no additional training. Across eight ASR backbones and six datasets at W4G128, \TARQ\ improves mean rare-Word Error Rate (rare-WER) without an aggregate-WER regression, achieves the lowest cross-corpus rare-WER swing among compared methods, and transfers to entity-rich benchmarks (ProfASR, ContextASR-Speech-En) without entity supervision.

The rapid adoption of large language models (LLMs) has shifted a substantial portion of inference workloads into throughput-oriented offline regimes, where fully utilizing GPU compute requires large batch sizes. However, existing deployments face a structural tension. Data parallelism (DP) scales throughput well but replicates model weights, leaving limited GPU memory for key-value (KV) cache and constraining batch size. Model parallelism reduces per-device weights, but requires fine-grained synchronization that erodes DP's independence and scheduling flexibility. We present SiDP, a memory-efficient data-parallel paradigm for offline LLM inference that treats weights as a bandwidth-backed shared resource inside a DP group. Instead of storing the full model on every GPU, SiDP organizes weights as a distributed pool: each layer is owned by a single GPU, and other replicas access its weights on demand via two complementary execution modes: a Weight-as-a-Service (WaS) mode that streams remote weights over NVLink into a small cache in the large-batch regime, and a Compute-as-a-Service (CaS) mode that ships activations to owners in the small-batch tail. Evaluated on NVIDIA H20, H200, and B200 GPUs with Qwen3-32B, Qwen2.5-72B, and Llama-3.1-70B, SiDP increases usable KV capacity by up to 1.8x under the same configurations, and converts this into up to 1.5x higher end-to-end throughput over baselines (vLLM) for offline workloads.

Modern large language model (LLM) inference has progressively disaggregated to keep pace with growing model sizes and tight TTFT and TPOT service-level objectives: from chunked-prefill aggregation, to prefill-decode (P/D) disaggregation, and most recently to operator-level Attention-FFN Disaggregation (AFD). This trend is especially important for mixture-of-experts (MoE) models, where memory-bound attention, compute-intensive expert FFNs, and MoE dispatch/combine communication create distinct resource demands. AFD further exposes this heterogeneity by placing attention and MoE-FFN execution on separate GPU groups. Each level of disaggregation deepens the scheduling design space across workload characteristics, resource allocation, and interconnect topology, raising the central question: when does each level actually pay off? We systematically characterize this trade-off for MoE inference across realistic workloads spanning input/output sequence lengths, prefix-KV reuse, and per-user latency constraints. Using chunked-prefill and P/D disaggregation as baselines, we study the benefits and limits of AFD at scale through a framework that fuses on-device kernel measurements with high-fidelity network simulation. Under strict TTFT/TPOT SLOs, AFD sustains around 4k tokens/s of system throughput on DeepSeek-V3.2 across chat, coding, and agentic-coding workloads, where non-AFD deployments are infeasible. We distill concrete takeaways for jointly optimizing throughput and interactivity, including how to partition attention and FFN across GPUs as a function of workload and model architecture, providing design principles for current rack- and cluster-scale deployments as well as future disaggregated AI infrastructure.

Electroencephalography (EEG) is a critical, non-invasive method to monitor electrical brain activity. EEGs can span anywhere from a couple seconds to multiple hours, posing a major hurdle for existing deep learning methods due to two major factors: (1) existing EEG models are predominantly built upon the attention mechanism, incurring quadratic scaling as the sequence length increases, and (2) raw EEG signals must be processed in a sliding-window fashion due to fixed-length input requirements, preventing global understanding of the entire signal. To this extent, we propose CaMBRAIN - the first Causal, Mamba-based state space model (SSM) capable of real-time inference of EEG signals, arguing that bidirectional approaches are needlessly expensive given the causal, unidirectional nature of EEG. However, training such a model is non-trivial, as crucial EEG events can be extremely brief - within fractions of a second - yet separated by long intervals spanning minutes. Current EEG methods use self-supervised objectives that optimize for signal reconstruction, but these are not well suited for streaming SSMs; they fail to explicitly train the hidden state to retain the salient long-range context needed for streaming inference. We therefore introduce a multi-stage self-supervised training pipeline specifically tailored to encourage long-range memory retention and strong performance on EEG signals, while preserving the linear-time complexity of state space models. CaMBRAIN achieves state-of-the-art (SOTA) results across 3 different EEG datasets with >10x higher throughput than existing models, enabling the first model capable of long-range, continuous inference of variable-length EEG signals.

Diffusion Transformers achieve strong video generation quality, but the quadratic cost of full attention limits efficiency. We introduce OSP-Next, an efficient text-to-video generation model that integrates sparse attention, parallelism, quantization, and reinforcement learning. OSP-Next uses a hybrid full-sparse attention architecture, where the sparse component is implemented with Skiparse-2D Attention. This fixed-pattern mechanism applies token-wise and group-wise sparse attention along spatial dimensions, leveraging locality while maintaining native compatibility with FlashAttention kernels. Based on the local equivalence of rearrangement in Skiparse-2D Attention, we further propose Sparse Sequence Parallelism (SSP), which partitions subsequences across ranks and switches sparse patterns through a single All-to-All communication. Compared with Ulysses Sequence Parallelism (SP), SSP provides a native parallel strategy for sparse attention and reduces communication volume by 75%. OSP-Next also incorporates HiF8 quantization to enable stable joint training with 8-bit quantization and sparse fine-tuning, and applies Mix-GRPO post-training to improve the performance of the sparse model. Experiments show that OSP-Next achieves a VBench total score of 83.73%, surpassing the Wan2.1 baseline. Under the 5-second 720P and 5-second 768P settings, OSP-Next achieves up to 1.64$\times$ single-GPU speedup and over 1.52$\times$ eight-GPU speedup on NVIDIA H200 GPUs. In addition, with only a 0.4% drop in VBench total score, OSP-Next-HiF8 achieves 1.69$\times$ and 2.27$\times$ speedups under the two settings on a single Ascend 950PR, demonstrating the efficiency and performance of OSP-Next across hardware platforms.

Pretrained foundation models have become an important basis for end-to-end autonomous driving. In contrast to vision-language models pretrained primarily on static image-text pairs, video generative models capture temporal dynamics and motion priors that are naturally suited for driving. We present DriveWAM, a driving world-action model that adapts a pretrained video diffusion transformer into an autoregressive video-action policy. DriveWAM organizes video and action streams into a unified temporal token sequence and trains them under a joint flow-matching objective, preserving the pretrained video-generation architecture while adapting its large-scale video priors to action generation. To incorporate high-level scene understanding, we introduce scene-evolving driving guidance, where a frozen VLM produces chunk-specific semantic intent to guide video-action generation. To keep long-horizon rollout bounded, we further introduce selective KV memory, which maintains bounded modality-aware video and action memory pools through relevance-redundancy cache selection at inference time. Experiments on NAVSIM and the PhysicalAI-Autonomous-Vehicles benchmark show that DriveWAM achieves strong planning performance, and a data-scaling study from 4k to 100k driving clips further confirms the scaling potential of world-action modeling for end-to-end autonomous driving.

World models for interactive video generation have largely focused on single-agent settings, where future observations are generated from a single control signal. However, many generated environments require multi-agent interaction: multiple players, robots, or embodied agents act simultaneously within a shared space. Scaling world models to such settings requires a principled multi-agent design: agents should remain independently controllable, permutation-symmetric, and support efficient inference while maintaining consistency across time and perspectives. In this paper, we present our generative multi-agent world model for interactive simulation. It introduces Simplex Rotary Agent Encoding, a parameter-free extension of 3D RoPE that represents agents as vertices of a regular simplex in rotary angle space. This gives each agent a distinct phase while making all agents permutation-equivalent, enabling scalable agent identity without learned per-slot identities or a fixed agent ordering. To avoid dense all-to-all attention across agents, we further propose Sparse Hub Attention, where learnable hub tokens mediate token interaction across agents, reducing cross-agent attention cost from quadratic to linear in the number of agents. For real-time rollout, we distill a full-context diffusion teacher into a causal student that generates temporal blocks sequentially with KV caching, enabling action-responsive generation at 24 FPS. Experiments in multiplayer virtual environments show that our model improves video fidelity, action controllability, and inter-agent consistency over slot-based and dense-attention baselines, while generalizing from two to four players without additional training.

Safety evaluations of language models often treat serving configuration as fixed background infrastructure, but batch condition is an untested treatment variable whenever the same prompt may be evaluated alone, in a synchronized batch, or inside a continuous-batching scheduler. We synthesize four artifact-backed studies into a paired testing protocol: Study A combines local discovery, scorer-corrected adjudication, and true-batching confirmation; Study B tests cross-model generalization; Study C tests continuous-batch composition; and Study D runs a batch-invariant-kernel ablation. The local test finds safety-label changes more often than capability-label changes (0.51% vs. 0.14%), but adjudication of 63 candidate rows leaves only 17 genuine behavioral flips, implying a corrected full-set rate of 0.16%. The 15-model extension finds no detectable universal safety-over-capability skew: flips are near parity (0.94x), alignment type has no detectable association ($p=0.942$, $η^2=0.033$), and output instability is the strongest tested fragility screen ($r=0.909$, bootstrap 95% CI [0.65, 0.97]). In the targeted kernel ablation, standard vLLM reproduces 22/55 label flips on current score-flip candidates, while enabling VLLM_BATCH_INVARIANT=1 reduces the same test to 0/55 flips; the composition test separately finds no aggregate effect at 4.7pp sensitivity. The testing recommendation is exact-stack validation: evaluate refusal at the served batch setting, pair safety prompts with capability controls, and report low-rate directional flips separately from aggregate null effects.

Large language model (LLM) serving creates environmental impacts beyond carbon and water, including ecosystem damage through biodiversity-related pathways. We present BIRDS, a framework for Biodiversity Impact of Request-Driven LLM Serving. BIRDS defines request-level functional units, quantifies operational and embodied biodiversity impact, and introduces Quality-Normalized Biodiversity Impact (QNBI) to jointly analyze ecological impact and response quality. Across diverse workloads, models, GPUs, and regions, BIRDS reveals that biodiversity impact accumulates at scale and exposes actionable quality-aware serving tradeoffs.

Multi-agent LLM systems have become the dominant production workload, but the serving stack was not built for them. The agent framework above knows agent identities, role, schemas, and dispatch structure but never sees an engine-level event; the serving engine below sees every event but knows nothing about agents. A surprising number of cross-cutting policies depend on both: prefix caching, batch shaping, speculative execution, fairness, tool-result memoization, safety enforcement, and more. Each lives in the seam between the two layers and is currently solved by a one-off patch into one neighbor or the other. We argue this seam is best addressed by an architectural change rather than point fixes: insert a third tier, an agent runtime layer, between the framework and the engine, exposing four primitives (observe, score, predict, act) into which any agent-aware policy plugs, with agent identity as the shared coordinate. We map nine concrete policies onto the layer and validate the abstraction in depth on the one with the largest immediate serving-cost lever: KV caching across sessions, instantiated as CacheSage, which learns the per-workload agent transition matrix online and uses it for survival-based eviction and between-step prefetch. Preliminary results on five real multi-agent workloads show +13 to +37 pp cache hit-rate lift, 12% to 29% lower mean TTFT, and 6% to 14% higher throughput over an unmodified serving stack.

Modern retrieval-augmented generation(RAG) deployments increasingly rely on caching to reduce token cost and time-to-first-token(TTFT). Prefix-level KV reuse is now standard in serving stacks such as vLLM, and chunk-level and position-independent reuse have been pushed further by recent systems(RAGCache, TurboRAG, CacheBlend, EPIC, ContextPilot, PCR, LMCache). Output-level semantic answer caches, by contrast, remain fragile: similar prompts can map to different correct answers, retrieved evidence drifts as the corpus is updated, and adversarial collision attacks have been shown to hijack cached responses. We argue that the right framing for cached answer reuse is not how to reuse faster but when reuse is safe. We propose GroundedCache, an evidence-validated cache router that admits a cached answer only when 4 cheap gates simultaneously hold: query similarity, retrieved-evidence overlap, source-version validity, and lexical (or judge-based) support of the cached answer by the freshly retrieved evidence. We build a six-regime workload that stress-tests cache safety rather than only hit rate, and introduce an operator-facing metric, the unsafe-served rate (USR), fraction of all queries that received a wrong cached answer. Across 2 datasets and 12,000 real-LLM generations(Qwen2.5-7B-Instruct on vLLM with Automatic Prefix Caching), GroundedCache drives USR to 0.0% on every HotpotQA regime(vs. 15-35% under naive caching) and to 1.5% on mtRAG document drift(vs. 51.5%), a 34x reduction on the design-point adversarial regime and 3-10x reductions across the other mtRAG regimes, while end-to-end p50 latency stays within 1.04-1.07x of a no-cache RAG baseline. A per-gate ablation isolates the lexical support gate as the load-bearing safety mechanism on both datasets, with the remaining gates providing defense-in-depth at near-zero cost. We release the implementation, workload, and evaluation harness.

We propose Hurwitz Quaternion Multiplicative Quantization (HQMQ), a calibration-free method for KV cache compression of large language models. HQMQ treats each 4-element chunk of K or V as a quaternion and quantizes its unit direction to the product $q_p \cdot q_s$, where $q_p$ ranges over the 24-element Hurwitz group $2T$ (the 24 vertices of the 24-cell on $S^3$, pairwise angle $60^\circ$) and $q_s$ ranges over a per-(layer, head) secondary codebook of $S$ random unit quaternions. The multiplicative composition yields $24S$ effective codewords at $S$ stored parameters; random initialization suffices because left-multiplication is an $S^3$ isometry, so seeded codebooks vary in end-task ppl by $<1.5%$. A per-batch median-multiplier outlier extraction step ($C{=}3$, no calibration) handles modern outlier-heavy architectures. We evaluate on five modern open models: Mistral-7B (dense MHA), Llama-3-8B and Qwen2.5-7B and Qwen3-8B (dense GQA), and gpt-oss-20b (sparse MoE). On Mistral-7B and Qwen3-8B, HQMQ matches fp16 within $0.02$--$0.03$ ppl points at $\sim$5 bits. On Qwen2.5-7B and Qwen3-8B, where naive int4 collapses to $10^4{+}$ ppl, HQMQ + Med3$\times$ recovers fp16 quality within $0.02$--$0.10$ ppl points at $\sim$5 bits. HQMQ Pareto-dominates naive int by $3$--$1900\times$ at matched bits across all five models, and downstream zero-shot accuracy matches fp16 at $3.79$ bits on Mistral. Against the strongest calibrated KV-quantization baseline, HQMQ at $3.79$ bits matches KIVI-4 ($\sim 4.5$ bits) within ${\sim}1$ pt on CoQA, $0.6$ pts on TruthfulQA, and $2.3$ pts on GSM8K, at $16%$ fewer bits and without a calibration pass. At the storage level, HQMQ delivers up to $5.05\times$ KV compression, shrinking a Llama-3-70B 128k-context cache from 43 GB to 8.5 GB.

Mobile graphical user interface (GUI) agents enable AI models to autonomously operate smartphones on behalf of users. However, most existing systems focus primarily on optimizing task accuracy and rely on cloud-hosted models for inference, which introduces privacy concerns and network-dependent latency. As a result, fully on-device deployment of mobile GUI agents remains underexplored. We propose MobileExplorer, a new framework that accelerates on-device inference for vision-based mobile GUI agents via online exploration. The key idea is to exploit the long per-step reasoning time of vision-language models (VLMs) by performing lightweight, parallel exploration of UI elements. During model inference, the agent proactively probes semantically relevant UI elements and records these exploration traces as structured memory. To ensure reliable execution in live mobile environments, we design a two-level rollback mechanism that robustly restores the initial UI state when a fast but naive backtracking strategy fails. The collected exploration traces are then summarized into concise contextual hints and injected into the prompt to enhance the subsequent reasoning step. We evaluate MobileExplorer on multiple off-the-shelf devices using the AndroidWorld benchmark, as well as newly designed, more complex tasks and dynamic on-device environments. MobileExplorer reduces the average number of reasoning steps and end-to-end latency by 23%, while maintaining or improving task success rates by up to 5%. A video demonstration of MobileExplorer performance in the real world is available at https://youtu.be/thK7MJmdlvM .

Fine-grained Mixture-of-Experts (MoE) models sparsely activate only a subset of experts per token, reducing activated computation while maintaining high model capacity. However, in memory-constrained inference scenarios, only a small set of experts can be cached. Experts not in the cache must be fetched from slow external storage (e.g., UFS), leading to frequent evictions and substantial I/O overhead. We propose ReMoE, a router fine-tuning framework designed to boost token-wise expert reuse. ReMoE biases the router toward recently selected experts, producing temporally stable routing that better matches cache locality constraints. By increasing short-horizon expert reuse, ReMoE reduces expert fetches from storage without adding inference-time computation. Experiments on DeepSeek and Qwen models show that ReMoE improves expert reuse by 26% while maintaining downstream task performance. Real-system evaluations further confirm these benefits, improving output throughput by 8.4% under vLLM GPU-CPU expert offloading and reducing TPOT by 43.6-49.8% under llama.cpp on Jetson Orin NX, corresponding to a 1.77-1.99$\times$ decode speedup across diverse workloads. Checkpoints and usage instructions are available at https://github.com/BUAA-OSCAR/ReMoE.