NeurIPS 2026 论文列表

Highlight detection in content-dense untrimmed videos, such as livestreams, is particularly challenging due to the prevalence of semantically similar moments and the application-driven demand of fitting a target duration budget. Existing methods typically rank segments independently based on semantic relevance, struggling to discriminate fine-grained quality among similar moments and allocating the budget to redundant selections. We therefore reformulate highlight detection as jointly selecting high-quality and complementary segments under a duration budget. To study this formulation, we introduce QB-Highlights, a large-scale benchmark of 50,000 e-commerce livestream videos with densely occurring query-relevant events and subtle quality variations. QB-Highlights is organized around two complementary dimensions: Segment Qualification, which assesses fine-grained per-segment quality; and Budgeted Composition, which evaluates set-level complementarity under a duration budget. It provides MLLM-assisted pseudo labels for training, dense human annotations for testing, and an evaluation protocol jointly measuring quality and diversity. We further propose a Qualification-then-Composition Reasoning (QCR) framework that performs structured deliberative reasoning to produce a highlight set matching the target duration. Extensive experiments show that QCR outperforms strong MLLM, moment retrieval, and highlight detection baselines on both segment quality and compositional diversity.

Large vision-language models (LVLMs) rely on increasingly long visual contexts, making the KV cache a major inference-time bottleneck. Existing KV cache compression methods typically use pre-decode signals such as prefill attention, visual redundancy, objectness, or layer-wise allocation, but these criteria do not directly indicate which visual tokens in the KV cache will be reused by future answer tokens. We introduce Q-ViK, a question-conditioned visual utility framework for LVLM KV cache eviction. During offline training, we run full-cache decoding and aggregate answer-to-visual attention over cached visual tokens, yielding a privileged *future-utility signal* that reflects the visual KV entries actually used during generation. We use this signal to train a lightweight question-conditioned scorer that predicts visual cache utility from prefill representations alone. At inference time, Q-ViK requires neither full-cache lookahead nor generated traces: it preserves textual KV entries and evicts low-utility visual KV entries with a single post-prefill scoring step. Experiments on seen and held-out multimodal benchmarks show that Q-ViK preserves task-relevant visual evidence in the KV cache more effectively than pre-decode saliency, especially under aggressive KV cache compression. Code will be available at [https://anonymous.4open.science/r/Q-ViK-1A75/README.md](https://anonymous.4open.science/r/Q-ViK-1A75/README.md).

Reinforcement Learning (RL) has empowered Multimodal Large Language Models (MLLMs) to achieve superior human preference alignment in Image Quality Assessment (IQA). However, existing RL-based IQA models typically rely on coarse-grained global views, failing to capture subtle local degradations in high-resolution scenarios. While emerging "Thinking with Images" paradigms enable multi-scale visual perception via zoom-in mechanisms, their direct adaptation to IQA induces spurious "cropping-implies-degradation'' biases and misinterprets natural depth-of-field as artifacts. To address these challenges, we propose Q-Probe, the first agentic IQA framework designed to scale IQA to high resolution via context-aware probing. First, we construct Vista-Bench, a pioneering benchmark tailored for fine-grained local degradation analysis in high-resolution IQA settings. Furthermore, we propose a three-stage training paradigm that progressively aligns the model with human preferences, while simultaneously eliminating causal bias through a novel context-aware cropping strategy. Extensive experiments demonstrate that Q-Probe achieves state-of-the-art performance in high-resolution settings while maintaining superior efficacy across resolution scales.

We present a novel theoretical framework, Q-MMR, for off-policy evaluation in finite-horizon MDPs. Q-MMR learns a set of scalar weights, one for each data point, such that the reweighted rewards approximates the expected return under the target policy. The weights are learned inductively in a top-down manner via a moment matching objective against a value-function discriminator class. Notably, and perhaps surprisingly, a data-dependent finite-sample guarantee for general function approximation can be established under only the realizability of $Q^\pi$, with a **dimension-free** bound---that is, the error does **not** depend on the statistical complexity of the function class. We also establish connections to several existing methods, such as linear FQE. Further theoretical analyses shed new light on the nature of coverage, a concept of fundamental importance to offline RL.

Video Large Language Models face severe computational and memory bottlenecks in long-video understanding. Existing visual token compression methods predominantly rely on static, query-agnostic redundancy elimination. Consequently, they lack the ability to dynamically preserve query-relevant details, failing to emulate the human "coarse-to-fine" cognitive process. To bridge this gap, we propose \textbf{Q-Focus}, a plug-and-play, two-stage visual token compression framework. Q-Focus first constructs a holistic semantic representation via a global compression stage. This is followed by a question-guided focusing stage driven by two novel components: the Adaptive Contrastive Module (ACM) and the Question Context Module (QCM). Specifically, ACM achieves adaptive query-visual alignment by treating the user queries as positive samples and dynamically sampling negative samples from the video's feature distribution. Complementarily, QCM leverages visual affinity matrices to diffuse semantic relevance from highly matched visual anchors to surrounding contextual regions, effectively capturing implicit yet critical details. Extensive experiments demonstrate that Q-Focus seamlessly integrates with diverse compression baselines. Remarkably, while retaining only 10\% of the original visual tokens, it achieves relative accuracy gains of 7.3\% and 6.6\% on LLaVA-OneVision and LLaVA-Video, respectively, and consistently boosts the performance across other state-of-the-art models such as Qwen3-VL-8B-Instruct.

Autoregressive video diffusion models (ARVDs) have emerged as a promising architecture for streaming video generation, paving the way for real-time interactive video generation and world modeling. Despite their potential, the substantial inference cost of ARVDs remains a major obstacle to practical deployment, making model quantization a natural direction for improving efficiency. However, quantization for ARVDs remains largely unexplored. Our empirical analysis shows that directly applying existing quantization schemes developed for standard diffusion transformers to ARVDs leads to suboptimal performance, revealing quantization behaviors that differ from those observed in bidirectional diffusion models. In this paper, we identify two critical challenges in quantizing ARVDs: (C1) Highly unbalanced frame-wise quantization sensitivity. Error accumulation during autoregressive generation can induce severely skewed quantization sensitivity across frames, following an exponential-like decay pattern. (C2) Prominent and heterogeneous outlier patterns in weights. Weight distributions exhibit pronounced outlier channels, whose patterns vary substantially across layer types and block depths. To address these issues, we propose Q-ARVD, a novel framework for accurate ARVD quantization. (S1) To tackle the highly unbalanced frame-wise sensitivity, Q-ARVD incorporates a final-quality aware frame-weighting mechanism into the quantization objective. (S2) To prevent heterogeneous outliers from degrading performance, Q-ARVD introduces an outlier-aware adaptive dual-scale quantization, which automatically detects the presence and quantity of outlier channels for an arbitrary layer, and isolates them to protect normal channels. Extensive experiments on state-of-the-art open-source ARVDs (i.e., self-forcing and causal-forcing) demonstrate the superiority of Q-ARVD. Practical deployment of INT8 model shows 1.30x speedup and 1.97x model size reduction.

Recent advances in sparse voxel-based VAEs have demonstrated remarkable capabilities in high-fidelity 3D autoencoding, yet their generative counterparts consistently lag behind their reconstruction performance. A fundamental cause of this gap lies in the \emph{brittle output representation} in sparse voxel decoders, which parameterize geometry through discrete topological decisions, such as intersection flags or occupancy signs. Small perturbations in these parameters, which are unavoidable under diffusion sampling, can be amplified into abrupt topological ruptures and grid-like artifacts. We argue that a generation-friendly representation should ensure that such perturbations induce only smooth geometric transitions. To this end, we propose Pygmalion, a generation-friendly autoencoding framework that reformulates sparse voxel decoding as a coupled SDF parameterization. To retain explicit mesh-level supervision within this parameterization, which is non-trivial as the SDF jointly governs both topology and geometry, we introduce hinge-based sign correction, case-aware geometry supervision, and rendering-based refinement, achieving high-fidelity reconstruction while preserving generative robustness. Built upon this generation-friendly representation, Pygmalion scales consistently across DiT model sizes and voxel resolutions. Extensive experiments demonstrate state-of-the-art reconstruction fidelity and high-quality 3D generation with complete surfaces, sharp features, and fine geometric details.

Understanding reflection remains a long-standing challenge in 3D reconstruction due to the entanglement of appearance and geometry under view-dependent reflections. In this work, we present the Pygmalion Effect in Vision, a novel framework that metaphorically “sculpts” reflective objects into clay-like forms through image-to-clay translation. We then introduce a dual-branch design in which a BRDF-based reflective branch and a clay-guided branch share the same Gaussian geometry but operate through independent rendering paths. The clay-guided branch, supervised by the synthesized clay images, provides reflection-free geometric guidance that complements the photometric supervision of the BRDF branch. Experiments on both synthetic and real datasets show consistent improvements in both geometry and photometric metrics. Beyond technical gains, our framework reveals that seeing by unshining, translating radiance into neutrality, can serve as a powerful inductive bias for reflective object geometry learning.

Biomolecular generators are often adapted with reward feedback to improve task-specific utility, but pushing utility alone can concentrate generation on a narrow family of candidates. Maintaining diversity is difficult because sample diversity is a set-level property. We introduce Supergroup Relative Policy Optimization (SGRPO), a flexible GRPO-style framework that directly constructs rewards from set-level diversity. For each condition, SGRPO samples a supergroup of candidate sets, compares their diversity under the same condition, and redistributes the group diversity reward to individual rollouts through leave-one-out diversity contributions before combining it with rollout-level utility. This design decouples SGRPO from a particular generator, utility reward, or diversity metric, and allows instantiation with different GRPO-style approaches. We evaluate SGRPO on de novo small-molecule design, pocket-based small-molecule design, and de novo protein design, instantiating it with both GRPO and Coupled-GRPO across autoregressive and discrete diffusion generators. Across decoding sweeps, SGRPO expands the utility-diversity Pareto frontier and achieves the best frontier-level metrics relative to pretrained generators, GRPO, and memory-assisted GRPO when applicable. Our analyses further show that direct set-level diversity rewards remain effective with small groups and help preserve broader generation-distribution coverage during post-training. The code is available at https://anonymous.4open.science/r/SGRPO/README.md.

Multi-agent systems (MASs) are increasingly deployed for complex tasks, yet their security against indirect prompt injection remains poorly understood. As frontier LLMs become more robust to injections with overtly malicious semantics, such as credential exfiltration, unauthorized transactions, or policy bypass, the residual attack surface shifts to task hijacking: an injection that silently redirects an MAS toward an attacker-chosen alternative task plausibly within the agent's domain, such as the wrong patient's record, a competitor's product, or a different CVE. There is no malicious verb to refuse, and existing benchmarks, designed around overtly malicious goals in single-agent settings, do not characterize this surface. We introduce PurpleAudit, an auditing framework that lets a practitioner stress-test their own MAS against task hijacking and derive a tailored system-prompt defense. The framework couples 22 baseline attacks and four prompt/tool-level defense baselines from the literature with a co-evolutionary purple-team procedure, in which red- and blue-team agents evolve stronger attacks and adaptive defenses by reasoning over each other's execution traces. Applied across frontier backends on ten MASs and a real-world MAS with live web access, PurpleAudit surfaces an MAS-specific attack pattern that frames the injection as a prerequisite for completing the user's task, reaching 31.26% ASR on Claude Opus 4.5 under the framework's auditing protocol, where 22 baseline attacks fail entirely. The co-evolved defense suppresses both baseline and evolved attacks at negligible utility cost, transfers across reasoning-capable backends, and operates at the prompt level with low deployment overhead.

We introduce the problem of best arm identification (BAI) with post-action context, a new BAI problem in a stochastic multi-armed bandit environment and the fixed-confidence setting. The problem addresses the scenarios in which the learner receives a \emph{post-action context} in addition to the reward after playing each action. This post-action context provides additional information that can significantly facilitate the decision process. We analyze two different types of the post-action context: (i) \textit{separator}, where the reward depends solely on the context, and (ii) \textit{non-separator}, where the reward depends on both the action and the context. For both cases, we derive instance-dependent lower bounds on the sample complexity and propose algorithms that asymptotically achieve the optimal sample complexity. For the separator setting, we propose a novel sampling rule called \textit{G-tracking}, which uses the geometry of the context space to directly track the contexts rather than the actions. For the non-separator setting, we do so by demonstrating that the Track-and-Stop algorithm can be extended to this setting. Moreover, in both settings, we theoretically and empirically show that algorithms that ignore the post-action context are sub-optimal. Finally, our empirical results showcase the advantage of our approaches compared to the state of the art.

Large vision model (LVM)-based gait recognition has achieved remarkable performance. However, current LVM-based methods largely rely on 2D image-plane representations and often neglect the underlying 3D structure of the human body. Without 3D structural guidance, 2D representations are sensitive to viewpoint variations. This sensitivity leads to unreliable gait features and limits generalizability in real-world 3D scenarios. To bridge this gap, we propose PuppetGait, a framework that aligns LVM features into a unified 3D canonical body space. The method conceptually manipulates the features like a puppet, adjusting them to a predefined pose and viewpoint. Comprehensive evaluations on CCPG, CCGR-MINI, CASIA-B*, and SUSTech1K show that PuppetGait maintains strong within-domain performance (96.54\% mean Rank-1 on CCPG) while substantially improving cross-domain generalization, reaching 23.22\% Rank-1 on CCPG-to-CCGR-MINI and 63.28\% mean Rank-1 on CCGR-MINI-to-CCPG. Code will be publicly available.

Large language models (LLMs) have demonstrated remarkable performance in video question answering (VQA). To improve responsiveness and answer quality, modern LLM-based VQA services pre-extract multiple LLM-compatible modalities from videos before response generation. However, the performance impact of modality combinations (MCs) and extraction knob settings has barely been studied. Our analysis shows that both factors substantially affect answer quality and Time-to-First-Token (TTFT) latency. Furthermore, the best configuration depends on both video content and query semantics, necessitating adaptive strategies. We propose Prysma, the first modality adaptation framework for LLM-based VQA. Prysma intelligently adapts the modality configurations to optimize answer quality, subject to a Service Level Objective (SLO) on average TTFT. It combines an offline modality orchestrator and an online semantic adapter. The offline orchestrator constructs a video knowledge base to warm-start tuning, employs a latency-aware configuration pruner to reduce search space, and runs a two-stage tuner to optimize both video- and query-level configurations. The online adapter selects configurations in real time based on query semantics and tuning history. Extensive evaluations demonstrate that Prysma improves answer quality by up to 82% over SOTA methods under the same SLO setting.

We study adversarial multi-armed bandits with and without delayed feedback under a **safety-aware** goal: achieving minimax-optimal worst-case regret while keeping nearly **constant** regret relative to a designated "safe" baseline policy. Existing approaches can balance this trade-off with immediate feedback for smooth comparators, but arbitrary delays can mistime transitions between **conservatism** and **exploration**, endangering the safety guarantee. To bridge this gap, we propose PRUDENT-BANKER, a novel algorithm that combines a delay-adapted variant of Online Mirror Descent with a modified phased-aggression mechanism. Its key technical contribution is a delay-calibrated restart threshold that rigorously accounts for the worst-case distortion induced by unobserved feedback and reliably detects comparator suboptimality. We also establish new lower bounds for safety-constrained adversarial delayed bandits, showing that the regret guarantees of PRUDENT-BANKER are unimprovable, up to logarithmic factors, under the baseline-safety requirement. To the best of our knowledge, PRUDENT-BANKER is the first algorithm to achieve the optimal safety–robustness trade-off: pseudo-regret $\tilde{O}(\sqrt{T}+\sqrt{D})$ together with $\tilde{O}(1)$ regret against the safe comparator, both with and without delays. Experiments across diverse delay distributions show that, unlike standard delay-robust baselines, PRUDENT-BANKER effectively balances safety and learning.

Noise search at inference time improves text-to-image diffusion quality but requires many multi-step teacher rollouts per query. Distilled few-step models are now routinely released for efficient inference, but their potential as proxies inside noise search is uncharacterised. We measure proxy--teacher rank agreement across three distillation regimes and three backbones spanning DiT, MMDiT, and UNet, and find a sharp asymmetry: eight closed-form image statistics preserve rank from the distilled proxy to the teacher (Spearman $\rho$ up to $0.80$), while learned semantic scorers decorrelate ($\rho \le 0.12$). Null controls localise the transfer to distillation training rather than few-step denoising or initial noise characteristics. We therefore reformulate noise search as a \emph{joint constraint} and propose ProxySearch: the proxy filters candidates by statistical match in $1$ step, the teacher argmaxes a user-chosen semantic reward over the survivors without a tunable weight trading semantic reward against the statistical target. \textsc{ProxySearch} matches or beats teacher-only argmax on compositional accuracy and human preference, while tightening the statistical target by $\sim 23\%$ in paired statistical $\ell_1$ relative to teacher-only argmax at matched compute on FLUX-$1024$.

Tracking the six-degree-of-freedom (6-DoF) pose of objects and surfaces from monocular video is a long-standing problem in computer vision. To tackle this problem, existing methods require inputs beyond the video itself—such as 3D models, depth maps, object masks, or task-specific learned features—and they struggle with textureless, transparent, reflective, or deformable surfaces. Here, we introduce ProxyPose, which recasts 6-DoF pose tracking as video-to-video translation. Given only a video and a single marked pixel in the first frame, a fine-tuned video diffusion model translates the input into a "proxy video"—a synthetic video depicting a colored polyhedron undergoing the same local rigid-body motion as the surface region at the marked pixel. Because the proxy's geometry and appearance are known by construction, recovering its full 6-DoF trajectory reduces to classical pose estimation with off-the-shelf solvers. This formulation leverages large-scale video pre-training to absorb the hardest aspects of pose tracking—handling challenging materials, occlusions, and deformations—into the translation step, while operating at the pixel level with no assumptions about object identity, boundaries, or global rigidity. ProxyPose achieves state-of-the-art 6-DoF pose tracking accuracy without the additional inputs required by competing methods and after fine-tuning the video model only on synthetic data. We further demonstrate that ProxyPose extends to face tracking, camera pose estimation, and challenging in-the-wild scenes that are beyond the reach of existing approaches. Video results are available on the [Supplemental Webpage](https://proxypose.github.io/).

Shapley and Banzhaf interactions capture the complex dynamics inherent in modern machine learning applications. However, current estimators for these higher-order interactions trade off between speed and accuracy. To overcome this limitation, we introduce ProxySHAP. ProxySHAP reconciles the high sample efficiency of tree-based proxy models with a principled path to consistency via residual correction. On a theoretical level, we derive a polynomial-time generalization of interventional TreeSHAP to compute exact interaction indices for tree ensembles, successfully bypassing exponential tree-depth dependencies in prior methods. Furthermore, we formally analyze the residual adjustment strategy, characterizing the specific conditions under which Maximum Sample Reuse (MSR) corrects proxy bias without its variance scaling exponentially with interaction size. Extensive benchmarking demonstrates that ProxySHAP sets a new state-of-the-art standard for approximation quality, including in large-scale applications with thousands of features. By achieving the lowest error in both small- and large-budget regimes, ProxySHAP significantly outperforms the prior best estimators ProxySPEX and KernelSHAP-IQ, while also delivering superior performance on downstream explainability tasks.

Estimating long-term treatment effects is essential for scientific and industrial evaluation, yet the limited follow-up duration of randomized controlled trials (RCTs) poses a significant challenge. Even the growing use of open-label extensions fails to resolve this issue, as such designs systematically censor long-term control trajectories. While supplementing RCTs with real-world evidence is a common strategy, existing methods often rely on restrictive exchangeability assumptions that fail under latent confounding. We propose a novel proximal framework to integrate long-term real-world evidence into RCTs with open-label extensions. Our approach successfully identifies long-term effects even in the presence of time-varying latent confounding effects and inter-study outcome drift. We provide a doubly robust locally efficient estimator that provides further resilience to model misspecification. Empirical results demonstrate that our framework maintains validity in complex data-generating processes that compromise standard approaches.

We study representation learning in low-rank Constrained Markov Decision Processes (CMDPs), where both value and reward functions are approximated by a set of unknown representation vectors, and the transition dynamics admit a low-rank factorization. The objective is to maximize expected cumulative rewards while complying with constraints on the expected cumulative utility. To achieve this, we propose \texttt{MFRLC} (Model-Free Representation Learning for Low-Rank CMDP), a model-free algorithm that uses a primal-dual scheme to effectively balance reward regret and constraint violations. To the best of our knowledge, \texttt{MFRLC} is the first representation-learning method for low-rank CMDPs that combines the Least-Squares Value Iteration with Upper Confidence Bound (LSVI-UCB) and primal-dual techniques. \texttt{MFRLC} further enhances value function estimation with bonuses for function approximation and uniquely learns the mapping from representations to value functions directly through representation learning. We prove that \texttt{MFRLC} achieves both regret and cumulative constraint violation of order $\widetilde O(H^3 d^2 K^{3/4}|\mathcal{A}|^{3/2}/\gamma)$, making it provably sample-efficient and highly adaptable to complex environments due to the reliance on function approximation.