NeurIPS 2026 论文列表

We investigate the problem of scaling safe reinforcement learning to massively parallel training regimes, where a large number of environments are executed with short rollout horizons. Although this setting improves data throughput and wall-clock efficiency, it creates a mismatch with constrained Markov decision processes, whose safety constraints are defined over full-episode cost returns. We address this mismatch by analyzing staggered environment resets as a phase-mixture distribution over short trajectory segments. This analysis characterizes the bias induced by phase coverage and motivates a phase-aggregated cost estimator that reconstructs episode-level cost estimates from staggered short rollout batches. We further incorporate finite-sample uncertainty and distribution mismatch into a constraint-tightening framework, yielding a scalable Safe RL method compatible with on-policy optimization.Using an MJX-based implementation of Safety Gym, our experiments show that the proposed method achieves performance comparable to non-massively parallel implementations while substantially reducing wall-clock training time

Diffusion and rectified flow (RF) models generate high-fidelity images and videos, but their iterative velocity-field evaluations are computationally expensive. Existing caching methods accelerate sampling by skipping timesteps, yet their coarse approximations introduce accumulated errors over long skip intervals and degrade quality under aggressive acceleration. We propose \textbf{TACache (Trajectory-Aware Cache)}, a training-free acceleration framework following a \textbf{skip-then-compensate} paradigm. TACache performs an orthogonal decomposition of discrete velocity acceleration along the RF trajectory into parallel and orthogonal components, isolating the magnitude and directional sources of per-step approximation error. The framework operates in two stages: \emph{offline}, cumulative variation thresholds on the magnitude and direction indicators yield the skipping schedule and bound how far each skipped interval may extend; \emph{online}, at each skipped step the offline statistics are combined with the sample's historical orthogonal direction to reconstruct the skipped velocity without additional forward passes. Experiments on BAGEL, FLUX.1-dev, and Wan2.1-1.3B show that TACache achieves up to $4.14\times$ speedup on text-to-image generation and $2.11\times$ speedup on text-to-video generation, with consistent improvements over prior cache-based methods on all reference-based fidelity metrics. Code will be released soon.

Neural network training can be viewed as a discrete-time dynamical system, suggesting that its optimisation trajectory can be learned and extrapolated to reduce the cost of repeated backpropagation. Motivated by Koopman operator theory, which represents nonlinear dynamics as linear evolution in a lifted space, we propose a training acceleration method based on augmented Koopman dynamics. Instead of approximating the evolution of parameters alone, our method lifts the training state to include both model parameters and optimiser-dependent internal states, enabling the learned operator to better capture the dynamics of modern optimisers with momentum or adaptive gradient statistics. We estimate a strided, multi-step Koopman operator from training snapshots and use matrix-vector multiplications to extrapolate future parameters and optimiser states, effectively bypassing an adaptive number of backpropagation steps. This formulation suppresses high-frequency stochastic fluctuations induced by mini-batch training and reduces the storage required for operator estimation. To improve robustness, we further introduce a safeguarding mechanism that prevents Koopman-extrapolated states from degrading performance relative to the most recent backpropagation iterate. The method can be integrated with standard optimisers, including Adam, AdamW, SGD with momentum, Adadelta, and other commonly used variants. Experiments across multiple optimisers show that the proposed approach significantly reduces both the number of backpropagation steps and the wall-clock training time required to reach a target accuracy, while maintaining, and in many cases improving, final model performance.

Long-context inference is increasingly important for privacy-sensitive local LLM deployment. In this setting, the prefill stage dominates Time-to-First-Token (TTFT), yet existing acceleration methods struggle to balance speed and quality. Sparse-attention methods accelerate only the attention module, leaving Feed-Forward Network (FFN) costs unchanged, while auxiliary-model-based compression often compromises quality under tight memory budgets. Recent layer-wise token pruning methods reduce both attention and FFN computation, but they typically rely on heuristic pruning schedules or token-recovery mechanisms that introduce decoding overhead. We propose FastPrefill, a recovery-free layer-wise token pruning system for efficient local long-prompt inference. FastPrefill employs an offline data-driven optimizer to determine layer-wise pruning schedules that maximize fidelity under a latency constraint, and executes these schedules via an algorithm–system co-design supporting head-specific sparse attention at runtime, where each head dynamically attends to its own selected key/value tokens. Evaluations on LongBench and RULER show that FastPrefill maintains comparable inference quality while achieving up to 2.13x TTFT speedup and 3.03x end-to-end speedup over state-of-the-art baselines.

Pre-training Large Language Models requires immense computational resources, making optimizer efficiency essential. The optimization landscape is highly anisotropic, with loss reduction driven predominantly by progress along flat directions. While matrix-based optimizers such as Muon and SOAP use fine-grained curvature information to outperform AdamW, their updates tend toward isotropy, which is relatively conservative along flat directions yet potentially aggressive along sharp ones. To address this limitation, we first establish a unified Riemannian Ordinary Differential Equation (ODE) framework that clarifies the mechanism of adaptive algorithms: the preconditioner induces a Riemannian geometry that mitigates ill-conditioning, while momentum serves as a Riemannian damping term that promotes convergence. Guided by these insights, we propose LITE, a generalized acceleration strategy that enhances training dynamics by applying larger Hessian damping coefficients and learning rates along flat trajectories. Extensive experiments demonstrate that LITE significantly accelerates both Muon and SOAP across diverse architectures (Dense, MoE), parameter scales (130M to 1.3B), datasets (C4, Pile), and learning-rate schedules (cosine, wsd). Theoretical analysis confirms that LITE facilitates faster convergence along flat directions in anisotropic landscapes, providing a principled approach to efficient LLM pre-training.

Diffusion Language Models (DLMs) exhibit strong parallel decoding capabilities by denoising multiple tokens in a single generation step. However, this parallelism comes with substantial computational overhead, as each step requires interactions with all suffix tokens. Existing methods typically reduce this cost by retaining only a local suffix window as a substitute for the full suffix. Despite their effectiveness, these methods overlook the structural heterogeneity across suffix regions and re-initialize suffix tokens with identical representations at each timestep. To this end, we propose a structured suffix modeling method for efficient DLM inference. Specifically, we divide the suffix into three regions, i.e., the local, middle, and tail regions, and retain different numbers of suffix tokens in each region according to their structural roles. Moreover, we incorporate the decoding results from the previous step into the suffix token representations at the current step, allowing them to carry evolving denoising information across generation steps. Notably, our method is training-free and orthogonal to several existing acceleration techniques, such as parallel decoding strategies and KV cache. Empirical results across multiple benchmarks on three DLMs demonstrate that our method can further accelerate DLM inference and, in most cases, also improve task performance. In particular, in long-sequence inference, our method achieves up to a \(72.16\times\) speedup when combined with other acceleration techniques.

Text-based image editing has recently been reinterpreted in large multimodal transformers as conditional generation, where source image tokens are concatenated with text and noise tokens. While effective, this design incurs substantial computational overhead in attention layers. To mitigate this inefficiency, we propose {\em TokenDrop}, an efficient editing framework that shift the computational burden from dense source-token conditioning to a lightweight regularized sampling update. By transferring the influence of source tokens into a closed-form sampling update, the method preserves source consistency with negligible cost. To support this regularized dynamics, we introduce a deviation-guided adaptive masking strategy that selectively drops redundant tokens while maintaining editing fidelity. Across FluxKontext and Qwen-Image-Edit, our training-free method achieves an average 22.4\% improvement in inference speed on PIEBench, while better preserving non-edited regions. The method delivers up to 1.8$\times$ speedup at 1024$^2$ resolution and 2$\times$ speedup at 2048$^2$ resolution. The code will also be released publicly.

Modern neural-network optimizers occupy a three-way design space: online adaptivity, architecture-aware geometry, and acceleration. We introduce AcceleGrad#, an accelerated linear-coupling template that couples an adaptive anchor with a norm-induced sharp, mirror, or linear-minimization geometry branch. The scalar-anchor specialization recovers the accelerated $O(1/K^2)$ rate for smooth convex objectives, while the diagonal-AdaGrad clipped-LMO training variant admits local $\mu$-Kurdyka-Lojasiewicz certificates: a generic conservative $O(K^{-1/4})$ rate and, in a SCION cross-entropy regime, a $O(K^{-1/3})$ last-iterate rate to a self-bounded mini-batch noise level. On image classification and language modeling, the training variant improves over Euclidean AcceleGrad and remains competitive with recent non-Euclidean optimizers. Using stochastic gradients directly in spectral oracles also reduces polar-computation time and optimizer overhead in practice.

Current LLM-for-GPU benchmarks largely measure CUDA fluency on isolated operators or short kernel-generation tasks, whereas practical GPU performance engineering often starts from trusted sequential CPU programs and requires correct, end-to-end acceleration. To fill this gap, we introduce \textsc{AccelEval}, a benchmark for whole-program CPU-to-CUDA acceleration. AccelEval contains 42 tasks derived from public repositories and established HPC/application suites across six domains, each equipped with deterministic correctness oracles and three input scales. Beyond pass/fail correctness, AccelEval reports end-to-end speedup over CPU baselines and a novel peak-attainment metric which normalizes each model’s speedup against the best correct solution. Using this framework, we evaluate eight frontier LLMs across multiple input scales, GPU architectures and sample rates. At pass@1 on NVIDIA H200 at medium scale, the strongest model passes 38/42 tasks at 71.2$\times$ geometric-mean speedup over the CPU baseline. On the subset of 11 tasks with human-written CUDA references, the best-of-8 LLM oracle reaches 0.75$\times$ of human CUDA performance. To make CUDA-optimization behavior interpretable and reusable, we decompose correct solutions into 43 CUDA acceleration patterns and show that feeding winner-derived pattern guidance back to models improves paired geometric-mean speedup by 1.78$\times$, compared with 1.18$\times$ for a generic-tips control. Together, these results establish AccelEval as a reproducible diagnostic platform for end-to-end LLM-assisted GPU acceleration. An anonymized open-source release is available at \url{https://anonymous.4open.science/r/AccelEval-2397}.

Enforcing state-wise safety constraints is critical for the application of reinforcement learning (RL) in real-world problems, such as autonomous driving and robot manipulation. However, existing safe RL methods often enforce state-wise constraints only in expectation, while hard state-wise guarantees typically require strong dynamics assumptions. The former can still allow rare but severe violations, while the latter is impractical in model-free stochastic systems. We instead target high-probability control of the maximum instantaneous cost along a trajectory. To accomplish this goal, we propose Absolute State-wise Constrained Policy Optimization (ASCPO), a model-free policy search algorithm whose exact update provides a distribution-free high-probability certificate for state-wise constraint satisfaction in stochastic systems. The guarantee holds under finite variance and does not require Gaussian, unimodal, or light-tailed cost distributions. We demonstrate the effectiveness of our approach by training neural network policies for extensive robot locomotion tasks, where the agent must adhere to various state-wise safety constraints. Our results show that ASCPO most consistently reduces state-wise safety violations while maintaining competitive reward across challenging continuous control tasks.

Designing antibodies that recognize a target antigen while avoiding non-cognate antigens remains a central challenge in computational antibody engineering. Existing antigen-conditioned generative models can jointly design antibody sequences and structures, but they are typically trained to match structural data distributions and lack an explicit alignment mechanism for antigen-specific preferences. We present AbSpecAlign, a plug-and-play specificity-reward alignment framework for antibody design. AbSpecAlign introduces a Bidirectional Multimodal Specificity Scorer (BiMSS), which integrates antibody and antigen sequence representations with SE(3)-equivariant structural embeddings to produce a contrastive reward over cognate versus non-cognate antigen pairs. We further introduce a GRPO-based post-training strategy that uses group-relative specificity rewards to align pre-trained antibody generators toward higher-scoring antigen-conditioned designs while maintaining structural regularity. To the best of our knowledge, AbSpecAlign is among the first frameworks to apply GRPO-style group-relative reward alignment to antigen-conditioned antibody sequence--structure co-design with an explicit off-target specificity reward. Experiments on RAbD CDR design benchmarks show that AbSpecAlign improves most sequence-recovery and structural-fidelity metrics when combined with diffusion-based antibody generators. Experiment results suggest that GRPO-based reward alignment with contrastive specificity signals provides an effective post-training mechanism for antigen-conditioned antibody generation, while prospective experimental validation remains necessary.

Complementarity Determining Regions (CDRs) are critical segments of an antibody that facilitate binding to specific antigens. Current computational methods for CDR design rely on reconstruction losses and do not jointly optimize binding energy, a crucial metric of antibody efficacy. Rather, binding energy optimization is performed via computationally expensive Online Reinforcement Learning (RL) pipelines that rely heavily on unreliable binding-energy estimators. In this paper, we propose AbFlowNet, a novel generative framework that integrates GFlowNet with Diffusion models. By framing each diffusion step as a state in the GFlowNet framework, AbFlowNet jointly optimizes standard diffusion losses and binding energy by directly incorporating energy signals into the training process, thereby unifying diffusion and reward optimization in a single procedure. Experimental results show that AbFlowNet outperforms the base diffusion model by 3.06% in amino acid recovery, 20.40% in geometric reconstruction (RMSD), and 3.60% in binding energy improvement ratio. ABFlowNet also decreases Top-1 total energy and binding energy errors by 24.8% and 38.1%, which is competitive with AbDPO's expensive online RL optimization without pseudo-labeling the test dataset or requiring computationally expensive synthetic CDR sampling.

Gigapixel object detection in High-Resolution Wide (HRW) images faces extreme spatial sparsity where targets often occupy less than 5% of the image area while computation is wasted on vast uninformative regions. Existing token selection methods require either complex evolutionary search or training additional learnable modules. We present AWP (Activation-based Window Pruning), which discovers that the mean of Feed-Forward Network (FFN) activations naturally encodes window importance, enabling parameter-free selection without additional modules. Leveraging this intrinsic signal, AWP performs window pruning with remarkable effectiveness. On the PANDA gigapixel dataset, AWP outperforms the previous state-of-the-art by +1.6% AP₅₀ with 26.2% FLOPs reduction, particularly excelling on small objects (+5.1% APₛ). When applied to Swin Transformer, AWP demonstrates strong generalization, retaining 97.4% of baseline performance with 50% windows pruned. Remarkably, AWP enables zero-shot application by directly replacing learned selection modules, improving baseline by +0.6% AP₅₀ without retraining. Our work reveals that window-based vision transformers inherently encode spatial importance through FFN activations, offering a simple yet powerful alternative to existing selection mechanisms, especially for High-Resolution Wide images.

Self-distillation enables language models to learn in an on-policy fashion from their own trajectories by using the same model as both the teacher and student, with the teacher being conditioned on privileged information that is unavailable to the student. This information can take multiple forms or views, including solutions, demonstrations, or feedback; for example, for a math problem, one view might be a ground truth reasoning trace. Supervising the student via the privileged teacher enables fine-grained token-level feedback and removes misalignment caused by distillation from external policies with different semantic biases. However, self-distillation also creates a fundamental asymmetry: the teacher may encode view-specific information tied to the privileged information provided during training, which the student cannot access at inference time. Moreover, the best form of privileged information is often task-dependent, making it difficult to choose a single teacher view. In this work, we address both these challenges jointly by introducing AVSD (Adaptive-View Self-Distillation), a novel method of self-distillation with multiple privileged-information views, which reconstructs token-level supervision by separating stable cross-view consensus from view-specific residual support. The consensus signal provides a robust update direction, while the residual signal is selectively used to adjust the update magnitude. Together, these components form a gated aggregation mechanism that balances a conservative target (geometric mean) with a more permissive target (arithmetic mean) for aggregating teacher signals at the token level. Experiments on math competition benchmarks (AIME24, AIME25, and HMMT25) show that AVSD consistently outperforms single-view self-distillation baselines and GRPO, achieving 3.1\% Avg@8 gain over the strongest baseline on Qwen3-8B on average. Moreover, on code-generation benchmarks (Codeforces, LiveCodeBench) using Qwen3-8B, AVSD outperforms single-view self-distillation baselines by 2.9\% on average. More broadly, our analysis shows that AVSD provides better token-level learning signals than any single-view method by reconstructing supervision from multiple privileged teachers.

Multimodal Large Language Models have achieved remarkable progress in short-form audio-video understanding, yet long-form audio-video comprehension remains challenged by limited context windows and severe information redundancy. To address these bottlenecks, we propose AVOC, a framework for long-form audio-video understanding in Omni-modal Large Language Models. AVOC introduces a learnable token compression module between the modality encoders and the LLM backbone. We reframe multimodal token compression as a top-$K$ retrieval problem: given a fixed context budget, the module must retrieve a compact subset of tokens that best supports answering the user query. We draw inspiration from three classical Information Retrieval criteria for selecting informative units from a large candidate pool: relevance, importance, and diversity. AVOC instantiates each criterion as a tailored mechanism for audio-video understanding, and integrates them into a unified retrieval-style compression pipeline. Experiments show that AVOC achieves state-of-the-art performance on long-form audio-video benchmarks, surpassing the second-best model by 4.9 and 5.5 points in average accuracy on OmniVideoBench and LVOmniBench, respectively. Moreover, AVOC maintains robust performance on Audio-Video Needle-in-a-Haystack task at durations up to one hour.

Modern Vision-Language Models (VLMs) benefit from chain-of-thought prompting and test-time scaling, but these gains often come with prohibitive inference cost due to large visual contexts and long decoding chains. We view this cost through two coupled axes: Visual Context Scaling (VCS), which controls how much visual evidence is passed to the language model, and Visual Reasoning Scaling (VRS), which controls how much inference-time reasoning search is performed. Existing methods typically optimize one axis at a time, leaving the joint allocation of compute across these axes underexplored. We introduce Adaptive Visual Inference Scaling (AVIS), a lightweight policy that adapts both VCS and VRS per query. AVIS realizes VCS through Key Diversity Visual (KDV) pruning, a training-free $O(N)$ key-based rule for removing redundant visual tokens before prefilling, and realizes VRS through adaptive self-consistency, using a learned difficulty predictor to select the number of reasoning rollouts. AVIS is deployment-friendly and compatible with shared-prefill inference, where all rollouts reuse a single prefilling pass and KV cache. Across diverse image and video reasoning benchmarks, AVIS improves the accuracy--compute trade-off relative to VCS-only and VRS-only baselines, and remains effective on top of RL post-trained VLMs while keeping compute and latency low.

We present AVI-HT, an adaptive visual-IMU fusion approach for tracking 3D hand poses by jointly modeling the egocentric image with on-glove 6-DoF IMU signals. AVI-HT achieves significantly improved accuracy and availability, particularly in hand-object interaction (HOI) scenarios involving heavy visual occlusion. Two complementary ingredients underpin its success: (1) synchronized multi-modal training data pairing on-body vision-IMU sensor streams with ground-truth 3D hand poses from a motion-capture system, and (2) a cross-sensor deep attention mechanism that adaptively modulates the trust assigned to the vision and individual IMU sensors. To evaluate AVI-HT in real-world settings, we conduct extensive experiments on our DexGloveHOI dataset that consists of 100K+ pairwise vision-IMU samples with synchronized 3D annotated poses, in which users manipulate a variety of objects during daily tasks. We compare against multiple single- and multi-modal tracking approaches under two hand models (UmeTrack, MANO). The results show that AVI-HT reduces mean keypoint error by 16.1% and its wrist-aligned variant by 24.2% over the baselines. Ablation studies further reveal the per-finger contribution of IMU sensors across activity types, and the model's sensitivity to IMU noise and temporal misalignment in vision-IMU fusion.

Audio-video (AV) editing aims to modify audio and video content according to a target prompt. Unlike single-modality editing, AV editing requires models to infer a modality-selective edit scope from the prompt alone: determining not only what should change, but also which modality should be preserved. Faithfully evaluating such models therefore requires both (i) benchmarks that span diverse edit types and modality categories, and (ii) evaluation that is itself modality-aware and sample-specific. However, existing AV editing benchmarks provide limited coverage of edit types and modality combinations, while current evaluation systems are often modality-blind and sample-agnostic, making it difficult to assess whether models faithfully preserve the unintended modality. To address these gaps, we introduce AVENUE, Audio-Video EditiNg Understanding and Evaluation, comprising two contributions: (1) a benchmark of 1,291 source clips and 7,957 editing instructions across audio-targeted, video-targeted, and AV-coupled edit types, curated and human-verified from VGGSound; and (2) a sample-specific, modality-aware evaluation framework that specifies, for each sample, both the intended change and the content that must remain intact. We evaluate representative AV editing models spanning three editing paradigms — joint, sequential, and separate — providing the first systematic analysis of modality-selectivity across paradigms. Our findings reveal a fundamental open challenge: when editing one modality, existing models frequently induce unintended changes in the other, regardless of paradigm. AVENUE provides a benchmark and modality-aware evaluation framework to drive progress toward more controllable AV editing models.

Memory expertise is a learned skill: knowing what to encode, when to retrieve, and how to organize knowledge---a capacity known in cognitive science as metamemory. We bring this perspective to LLMs by treating memory management as a trainable skill. We promote file-system operations to first-class memory actions alongside task actions, letting the model itself decide how to manage its memory. This memory skill improves along two axes: the structure that supports it (prompts, file schemas, action vocabulary), and the proficiency of the model exercising it. Both axes resist manual optimization: episodes in long-horizon tasks run for thousands of steps, and a single memory mistake can hide long before it surfaces, making human review of full trajectories impractical. We introduce AUTOMEM, a framework that automates both axes. In the first loop, a strong LLM reviews complete agent trajectories and iteratively revises the memory structure that shapes how the agent interacts with its memory files. In the second loop, the agent's own good memory decisions are identified from many episodes and used as training signal to sharpen the model's memory proficiency directly. Across three procedurally generated long-horizon games (Crafter, MiniHack, and NetHack), optimizing memory alone---without modifying the model's task-action behavior---improved the base agent's performance 2x--4x, bringing a 32B open-weight model competitive with frontier systems such as Claude Opus 4.5 and Gemini 3.1 Pro. Our results show that memory management is an independently learnable skill, and a high-leverage objective yielding large gains on long-horizon tasks.