NeurIPS 2026 论文列表

Harmful-content detectors are increasingly deployed in safety-critical settings, but they remain vulnerable to subtle rhetorical reframings that preserve the underlying claim while altering its presentation. Existing adversarial robustness methods are typically token-level or prompt-driven, hindering mechanism-level attribution and targeted repair. We propose Adaptive Rhetorical Adversarial Optimization (ARAO), which operationalizes rhetorical theory as an axis-aligned strategy space with directional and interpretable controls for adversarial rewriting. A planner-rewriter pipeline converts strategy vectors into non-contradictory rewrites, enabling mechanism-level failure attribution and axis-targeted repair. Experiments on misinformation and extremist-rhetoric benchmarks show that ARAO improves robustness by 2.35 ROC-AUC points on average over the strongest baselines on attacked sets, while achieving the best performance on original content.

The emergence of specialized, domain-tuned Large Language Models (LLMs) has demonstrated that smaller models can achieve expert-level performance in specific tasks, while struggling in out-of-domain settings. Current ensemble methods to combine their complementary expertise primarily rely on iterative re-prompting or cross-model refinement. These approaches suffer from high computational costs and latency because they require repeated LLM inference calls. Furthermore, naive aggregation often leads to anchor corruption, in which noise propagated from weaker models degrades the performance of the most accurate expert. To address these challenges, we propose a framework that integrates model predictions at the semantic layer using a bipartite factor graph. In this architecture, individual LLMs are represented as variable nodes, while a set of check nodes assess their consistency based on diverse epistemic criteria. We develop a message-passing protocol inspired by error-recovery systems to resolve disagreements iteratively. Furthermore, we introduce an asymmetric damping mechanism that protects high-reliability anchor nodes from being overridden by the ensemble majority. Unlike existing methods, our approach operates entirely on output distributions and requires no additional LLM calls during the refinement phase. Evaluating on four benchmarks, including MMLU, MMLU-Pro, GPQA, and MedMCQA, our method demonstrates a 97% reduction in token usage and up to a six-fold decrease in API calls, reducing inference time from several minutes to mere milliseconds while consistently outperforming leading multi-agent baselines. These results suggest that graph-based belief propagation is a robust, high-speed, and scalable alternative to the current multi-agent LLM systems. The full pipeline and code will be made public.

Text-to-image prompt optimization is often treated as prompt rewriting: given a user prompt, a language model expands it with richer style, composition, and photographic details. While this paradigm can improve visual appeal, it may silently weaken the user's actual intent, especially when the prompt specifies fine-grained objects, attributes, and relations. We argue that faithful prompt optimization requires a different objective: improving perceptual quality while preserving the compositional structure that makes the prompt semantically correct. We propose \ours{}, a single-step reinforcement learning framework that transforms noisy scene-graph feedback into a learnable prompt-editing policy. Instead of freely rewriting the entire prompt, \ours{} first extracts a calibrated scene graph from the input and then performs conservative deletion, reordering, and insertion actions conditioned on both text and graph representations. This design exposes structured intent to the optimizer while restricting unnecessary prompt drift. To make object--relation feedback usable for policy learning, \ours{} replaces sparse thresholded grounding rewards with uncertainty-aware smooth rewards and estimates advantages with environment-aware grouping over multiple image samples from the stochastic text-to-image generator. Across prompt-optimization datasets and generation backbones, \ours{} improves the faithfulness--aesthetics trade-off, with particularly consistent gains in relational fidelity where generic rewriting and aesthetics-oriented optimization often fail. We further evaluate with both human preference and Gemini-VQA, showing that the improvements are not confined to the SG reward used during training. Appendix additionally compares against a high-budget iterative GPT-4o structural-feedback baseline to contextualize the efficiency of amortized policy learning. More broadly, our results suggest that structural constraints should not remain only post-hoc evaluators of generated images; when properly calibrated and stabilized, they can be internalized as training signals for reliable, intent-preserving prompt policies.

Detecting training data memorization in diffusion models is important for copyright protection and privacy auditing. We view memorization as an abnormally local concentration process: during reverse-time generation, a memorized instance acts as a point-like attractor whose probability basin is sharper and more position-sensitive than that of a generalized concept. Starting from the Fokker–Planck equation, we derive an exact evolution law for the score field $\partial_t \mathbf{s}_t$ and show that its leading geometric contribution in score-dominated local regimes is $\mathbf{H}_t \mathbf{s}_t$, linking temporal concentration dynamics to spatial curvature. We then show that standard sampling along a shared guided trajectory already exposes this curvature information through a complementary transport-curvature response, without requiring explicit Hessian computation. Motivated by this analysis, we propose the **Dynamical Singularity Metric (DSM)**, an on-trajectory detector that measures the pathwise score-evolution discrepancy between conditional and unconditional branches. DSM requires no additional backpropagation or network evaluations beyond sampling itself. Experiments on Stable Diffusion show that DSM matches or exceeds curvature-based baselines while being substantially cheaper and effective at the earliest reverse step.

Direct Preference Optimization (DPO) is widely used to reduce hallucination in Multimodal Large Language Models (MLLMs). Prior studies and our observations show that during DPO training, the log probabilities of chosen and rejected responses can decrease simultaneously, revealing the squeezing effect. For MLLMs, the redistributed probability mass may move toward responses driven by language priors rather than visual evidence, weakening the role of the image and amplifying hallucination. Existing squeezing-aware methods constrain rejected updates with $\mathcal{V}$-usable information or generation confidence, while methods with anchor terms regularize the DPO implicit reward of chosen responses. Neither explicitly grounds the control signal in visual evidence. We propose Visual Guided Direct Preference Optimization (VGDPO), which relaxes negative updates on rejected responses according to multimodal squeezing risk. VGDPO estimates this risk by combining token-level visual dependency with phrase-level hallucination localization, defining a hallucination ratio for each rejected response. We further apply a related reweighting principle to a visual contrastive objective, providing additional supervision for chosen responses under the original image. Experiments on four hallucination benchmarks and three MLLM backbones show that VGDPO improves DPO log probability dynamics and reduces multimodal hallucination while maintaining informative responses.

Knowledge distillation from large reasoning models (LRMs) often uses raw teacher trajectories as supervision. However, this practice creates a mismatch between the goal of distillation and the nature of teacher trajectories. Advanced LRMs typically improve their reasoning capabilities through reward-driven post-training, such as RLHF. This training optimizes them as strong solvers, rather than as teachers that produce reasoning processes smaller models can easily learn. Since smaller models have different "cognitive capacities" compared to their larger teachers, directly imitating raw teacher trajectories can sometimes be ineffective and often requires a substantial number of high-quality samples. To bridge this mismatch, we propose **C**ognition-**A**ligned **R**easoning **D**istillation (**CARD**), an RL-based framework that transforms solver trajectories into teaching trajectories aligned with the student’s cognition. Avoiding expensive resampling, our framework directly transforms existing trajectories. We optimize CARD with two reward signals: (1) a Distribution Alignment Reward that matches the student’s predictive distribution, and (2) a Logical Coherence Reward that reduces misleading failed reasoning steps. Comprehensive evaluations across multiple student models and settings demonstrate that CARD outperforms distillation from raw teacher trajectories in in-domain settings while achieving substantial gains in out-of-distribution performance.

Multimodal Large Language Models (MLLMs) can listen and see, but how do audio and visual signals actually travel through the network to shape an answer? Despite their growing role in research and real-world applications, the internal pathways through which audio and visual tokens influence the final prediction remain poorly understood. In this study, we examine audio-visual information flow inside Audio-Visual Large Language Models (AVLLMs), tracing how AVLLMs route, utilize, and integrate audio and visual information across two input configurations, audio-visual video and multiple interleaved audio-visual items. We find that for audio-visual video, AVLLMs follow the sequential information flow pathway established for VLMs and VideoLLMs, with audio and visual contribution flowing along this pathway in proportion to the task's reliance on each modality. In settings with multiple interleaved audio-visual items, this routing shifts to different parallel streams. Furthermore, we demonstrate that audio-visual and other token types can be discarded once their information is transferred to LLM, with minimal impact on the model's prediction or even slight improvement, generalizing across multiple tasks and datasets, enabling more efficient inference. These findings hold across multiple models and scales, Qwen2.5-Omni and Video-SALMONN2 Plus at 3B and 7B scales, leading to hypotheses on why these flow structures emerge. Together, these results deliver the first coherent picture of how AVLLMs orchestrate sound and sight inside the network and lay the groundwork for the next wave of interpretability, design, and efficiency advances in audio-visual and broader MLLMs.

Although Large Vision-Language Models (LVLMs) have achieved remarkable progress, complex visual reasoning remains challenging. Existing approaches suffer from fundamental limitations: *Thinking about Image* mainly performs reasoning in text space, *Thinking with Image* relies on external tools to introduce additional visual evidence, and *Latent Visual Reasoning*, despite moving reasoning into latent space, still focuses largely on static visual modeling and lacks explicit characterization of temporal dynamics. We therefore propose **Dynamic Latent Visual Reasoning (DLVR)**, a training framework that enables LVLMs to reason about dynamics in continuous latent space from only a static image. We introduce Mahalanobis Novelty Token Selection and Novelty-Adaptive Temporal Quantization to construct dynamic latent supervision from open-source datasets, building DLVR-SFT-100K and DLVR-RL-4K. We further develop a two-stage SFT pipeline that first builds temporal grounding over explicit dynamic processes and then teaches the model to encode dynamic semantics and temporal structure into latent tokens. Finally, we propose Contrastive Dynamic Latent Policy Optimization, which encourages latent trajectories to align with real dynamics while moving away from counterfactual ones. DLVR empirically improves both dynamic-centric and general visual reasoning, achieving an average gain of 10.82% on BabyVision and a 10.50% improvement on HRBench4K, demonstrating the promise of dynamic latent reasoning for LVLMs.

Scaling robot demonstrations does not guarantee better imitation: redundant trajectories dilute learning, while rare, imperfect segments may encode critical transitions. Prior curation methods overlook such dependencies by scoring samples independently at the trajectory or segment level. We reframe robot demonstration curation as a coverage-aware sequential selection problem and propose PROSE, which selects data by maximizing marginal utility under a fixed budget. PROSE unifies three signals: (i) influence on closed-loop return, (ii) kinematic reliability, and (iii) coverage gain measuring novelty with respect to the retained set, allowing full trajectories and fine-grained segments jointly participate in scoring and filtering. Across three Robomimic simulation tasks and three real Franka tasks, PROSE achieves the best success rate on every task we evaluate against three state-of-the-art curators (CUPID, SCIZOR, Demo-SCORE), with average closed-loop gains of +15\% in simulation and +35\% in real-world. Ablations attribute the gain primarily to the coverage step.

Modern optimizers, like Muon, impose matrix-wise geometry constraints on their updates. These matrix-wise constraints can be unified under Linear Minimization Oracle (LMO) theory. However, all current methods impose fixed LMO geometries for the update rules, chosen by-design or empirically, which are not necessarily optimal according to the problem's geometry. We introduce a novel efficient data-driven criterion for dynamically choosing proxy-optimal update LMO geometries on individual Deep Neural Network layers. Derived in closed form from gradient and activation statistics using a single-step random feature regression surrogate model, our criterion navigates a design space interpolating from SGD to Muon updates. Moreover, integrating parameter-wise preconditioning allows our framework to recover SGD, Muon, Adam, and MuAdam as specific extrema. To make this adaptive approach scalable, we pair it with efficient computational strategies, achieving only a $\sim 3$% runtime overhead on highly optimized baselines. As a proof of concept, we show that this data-driven optimizer matches or exceeds the performance of the best performing optimizer between Muon and AdamW across three different training scenarios. Ultimately, this work provides evidence that LMO geometry can be successfully and efficiently adapted from runtime data, opening a new pathway for optimizer design beyond static geometries.

Recent work has shown that large language models form linear representations of emotion concepts that causally influence alignment-relevant behavior, including blackmail, reward hacking, and sycophancy. These findings raise a natural question: can such representations be used not only to characterize model behavior in controlled evaluations, but also as practical tools for monitoring and guiding models during deployment? We explore this possibility, investigating how emotion-vector activations can serve as interpretable signals of a model's internal state over the course of a conversation or agentic task. We discuss how these signals might support debugging, early detection of drift toward misaligned behavior, and lightweight steering interventions to nudge the model back toward desired behavior. Beyond the Assistant, the same representations apply to other characters in the model's context — including the user — suggesting a broader family of applications in which internal representations of emotional state inform how systems respond. We outline the opportunities and limitations of this approach and highlight open questions around validation, robustness, and responsible use.

Reinforcement learning from verifiable rewards (RLVR) has shown strong promise for LLM reasoning, but typical outcome-based RLVR methods remain inefficient on hard problems. Correct final-answer rollouts are rare, and standard sample-level credit assignment fails to leverage partial reasoning progress embedded in unsuccessful attempts. To address this problem, we introduce **SCRL** (**S**ubproblem **C**urriculum **R**einforcement **L**earning), a curriculum reinforcement learning framework built on verifiable subproblems derived from reasoning chains. Given a reference solution, SCRL derives a series of verifiable subproblems and constructs a subproblem curriculum, with the final subproblem fixed as the original problem. This converts partial progress on hard problems into verifiable learning signals. Algorithmically, we propose *subproblem-level normalization*, a training technique based on RLVR that normalizes rewards independently at each subproblem position within the rollout group. By assigning the resulting advantages to the corresponding answer spans, we enable finer-grained credit assignment without external rubrics or reward models. Our theoretical analysis shows that this subproblem curriculum makes hard problems more learnable by lifting them out of gradient dead zones, with larger relative gains as the original problem becomes harder. Across seven mathematical reasoning benchmarks, SCRL outperforms strong curriculum-learning baselines, yielding +4.1 and +1.9 average-point gains compared to GRPO on Qwen3-4B-Base and Qwen3-14B-Base respectively. On three hard benchmarks (AIME24, AIME25, and IMO-Bench), SCRL further yields point gains of +3.7 in pass@$1$ and +4.6 in pass@$64$ on Qwen3-4B-Base, suggesting improved exploration on hard reasoning problems. Further ablations show that the proposed credit assignment is effective and that the gains do not require highly curated subproblems or strong external generators, highlighting SCRL as a practical curriculum framework that enables fine-grained credit assignment for LLM reasoning.

Language agents increasingly improve by reusing \emph{skills}---structured procedural artifacts distilled from past experience. In particular, \emph{domain-level} and \emph{model-generated} skills are especially promising. They offer fast adaptation within a domain by encoding domain-specific recurring procedures, and they scale beyond labor-intensive hand-crafting. However, while extraction methods continue to proliferate, understanding remains limited, with no comprehensive study spanning the full skill lifecycle---\textbf{experience generation}, \textbf{skill extraction}, and \textbf{skill consumption}---to ask whether such skills actually work, when they work, and what makes them succeed or fail. To close this gap, we build a utility-grounded evaluation framework that provides systematic experimental results across extractors and target agents, covering five diverse agentic task domains. We find that model-generated skills are beneficial on average but exhibit non-trivial negative transfer, and that neither extractors nor targets behave uniformly. A model can be a strong extractor yet a weak consumer, or vice versa, with skill utility independent of model scale or baseline task strength. To explain these patterns, we then dissect each lifecycle stage in depth, analyzing how experience composition shapes skill quality, what properties characterize useful skills, and how the same skill transfers across different consumers. Finally, we translate these findings into a concrete \emph{meta-skill} that guides skill extraction toward the features tied to actual utility, which consistently improves skill quality across domains and substantially reduces negative transfer.

Semi-Supervised Time Series Classification (SS-TSC) with Deep Neural Networks (DNNs) achieves strong accuracy by leveraging unlabeled data, yet the resulting models remain black boxes that offer no insight into which features drive each prediction. Existing post-hoc explanation methods for TSC can identify important features but provide no guarantee that the model genuinely relies on them, while ante-hoc approaches impose architectural constraints that limit flexibility in existing model usages. Neither setting explore the semi-supervised regime, where limited labels can affect explainability. We propose \textbf{Sensory Gating}, the first SS-TSC framework that simultaneously bridges post-hoc attribution and ante-hoc explanation through a four-stage pipeline: (i) a base classifier is trained with our proposed Forecasting Joint-Embedding Predictive Architecture (F-JEPA) auxiliary objective; (ii) post-hoc attribution maps are converted into per-instance binary masks via greedy selection of sufficient features; (iii) a lightweight \textbf{Sensory Gate} amortizes these masks through Binary Cross Entropy (BCE) supervision with data-driven threshold calibration; and (iv) a knowledge distillation stage to provide an explainable SS-TSC student model that operates on gated (masked) input, while consistently reproduces the original black box teacher (F-JEPA)'s predictions. We validate our framework on eight datasets across four label ratios with five seeds per configuration. F-JEPA achieves better performance compared to recent state-of-the-art (SOTA) semi-supervised TSC methods, while the sensory gate reduces the fraction of features the classifier requires to 20--45\% with competitive accuracy. On MIT-BIH, where ground-truth saliency annotations are available, the gate's selected features align with domain-relevant regions while classification performance remains competitive with the ungated baseline at 10--20\% kept ratio.

Persistence diagrams are common representations in topological data analysis, yet they lack both the vector space structure required by machine learning algorithms and a principled framework for statistical comparison. We introduce STRAND (Survival Topological Representation ANalysis of Diagrams), which treats (collections of) PDs as survival data: each topological feature with persistence value $p = d - b$ is a fully observed time-to-event, and the persistence survival function $S(t) = \mathbb{P}(p > t)$ is the central object for comparing diagrams. From this single representation we derive (i) a non-parametric two-sample test with calibrated Type~I error and high power from a small number of diagrams; (ii) interpretable effect sizes; and (iii) a 1-Wasserstein-stable feature vector for downstream machine learning. We validate calibration and power on synthetic manifolds with controlled topology, demonstrate competitive vectorisation across 14 graph and 3D point cloud benchmarks, and apply the method to study functional brain connectivity in fMRI/neuroscience data. To our knowledge, STRAND is the first method to provide hypothesis testing and vectorisation for persistence diagrams from a single coherent and interpretable representation.

Generating humorous memes is a challenging multimodal task that moves beyond direct image-to-caption supervision. It requires a nuanced reasoning over visual content, contextual cues, and subjective humor. To bridge this gap between visual perception and humorous punchline creation, we propose HUMOR, a novel framework that guides VLMs through hierarchical reasoning and aligns them with group-wise human preferences. First, HUMOR employs a hierarchical, multi-path Chain-of-Thought (CoT): the model begins by identifying a template-level intent, then explores diverse reasoning paths under different contexts, and finally anchors onto a high-quality, context-specific path. This CoT supervision, which traces back from ground-truth captions, enhances reasoning diversity. We further analyze that this multi-path exploration with anchoring maintains a high expected humor quality, under the practical condition that high-quality paths retain significant probability mass. Second, to capture subjective humor, we train a pairwise reward model that operates within groups of memes sharing the same template. Following established theory, this approach ensures a consistent and robust proxy for human preference, even with subjective and noisy labels. The reward model then enables a group-wise RL optimization, providing a theoretical guarantee for monotonic improvement within the trust region. Extensive experiments show that HUMOR empowers various VLMs with superior reasoning diversity, more reliable preference alignment, and higher overall meme quality. Beyond memes, our work presents a general paradigm for open-ended, human-aligned multimodal generation, where success is guided by comparative judgment within coherent output groups.

Supervised fine-tuning (SFT) on long teacher trajectories is the dominant method for instilling investigation and reasoning capabilities into open software-engineering (SWE) agents. Under SFT, every retained response is an imitation target, so the student inherits not only the trajectory's outcome but also any flaw in its intermediate steps, including ungrounded leaps and redundant loops. High-quality training data must therefore be jointly effective (each step is grounded and narrows the agent's epistemic gap to the correct fix) and efficient (each step is information-bearing rather than redundant or looping). Existing recipes filter or relabel teacher rollouts using only a binary terminal verifier, which does not directly target these axes and provides no supervision on instances where the teacher fails. Every real issue ships with a developer-authored reference patch $p^\star$ that implicitly testifies to the file paths, runtime behaviors, and conventions a fix presupposes, but the standard pipeline discards it. We propose P2T (Patches-to-Trajectories), which uses $p^\star$ as privileged information during curation, and frames trajectory construction as a bi-objective program over per-step effectiveness and trajectory length. A reverse phase distills $p^\star$ into a latent process graph $G^\star$ of contextual facts and solution milestones, encoding dense intermediate anchors in constructive ordering. A forward phase curate trajectories from blinded teacher continuations, scoring per-step progress against $G^\star$ under a leakage-blocking groundedness check and committing the shortest segments that retain effectiveness. Using only 1.8k curated SWE-Gym instances, P2T improves both axes simultaneously over outcome-filtered SFT and its tool-error-masking variant: on SWE-bench Verified, it lifts Pass@1 by up to +10.8 points while cutting per-instance inference cost by 15%, with consistent gains on SWE-bench Lite and across two teachers. A size-matched ablation and qualitative analysis further isolate per-trajectory quality from data scale.

Model-based reinforcement learning in partially observable environments requires inferring hidden states, learning latent dynamics, and planning under the resulting belief. Practical algorithms often solve one aspect of the above pipeline while relying on approximations for the others that are often poorly understood, while existing POMDP theory typically analyzes computationally inefficient theoretical algorithms far from modern deep RL practice. To fill this gap, we propose a practical end-to-end algorithm that approximately solves the hidden state filtering problem via sequential Monte Carlo (SMC), updates a latent dynamics model online, and performs approximate online planning for $d$ steps before bootstrapping leaves with a state-value critic (a QMDP approximation). We analyze each approximation and prove an end-to-end regret bound decomposing the error into model estimation error, particle filter error, Monte Carlo rollout error, and a structural bias term induced by the QMDP approximation controllable by the planning depth that we further analyze in various special cases. The resulting guarantee makes explicit how statistical sample size, particle budget, rollout budget, and planning depth affect performance. We instantiate our algorithm in an MPC-based deep RL framework, demonstrating its practicality on a variety of environments. Our results provide a practical and theoretically grounded route for bridging theory and practice within online decision-making in POMDPs.

Large Language Models (LLMs) demonstrate remarkable reasoning capabilities, yet the internal mechanisms driving these multi-step processes remain opaque. Existing mechanistic interpretability approaches often rely on superficial text-pattern co-occurrences or capture only short-term effects, struggling to trace the long-horizon, outcome-driven influence of internal components such as neurons or sparse features in multi-step reasoning. In this paper, we introduce Integrated Policy Gradient (IPG), a novel, training-free framework that brings policy-based attribution into mechanistic interpretability for LLM reasoning. Grounded in *outcome-oriented* and *sequential-influence-aware* principles, it backpropagates outcome-level, non-differentiable behavioral signals (e.g., reasoning outcome correctness) through entire inference trajectories, utilizing path integration to isolate outcome-relevant mediators at component level for reliable attribution. This provides a linear approximation of Causal Mediation Analysis (CMA) in reasoning. Empirical evaluations demonstrate that IPG achieves precise localization and reveals mechanistic insights by identifying sparse, concentrated subsets of internal components associated with specific reasoning behaviors. Our method also enables modulation of reasoning capabilities, offering a powerful and reliable framework for understanding LLM reasoning.