NeurIPS 2026 Papers

Large language models are increasingly used for mental health support, yet their conversational coherence alone does not ensure clinical appropriateness. Existing general-purpose safeguards often fail to distinguish therapeutic disclosures from genuine clinical crises, and progress on developing better ones is bottlenecked by a lack of clinically grounded evaluation resources. To address this gap, we introduce a risk taxonomy, developed in collaboration with PhD-level clinical psychologists, that identifies actionable harm (self-harm and harm to others) while preserving space for safe, non-crisis therapeutic content. We release MindGuard-testset, a dataset of multi-turn conversations annotated at the turn level by clinical experts, which is, to our knowledge, the first public benchmark with turn-level clinical risk labels for multi-turn mental health support. We also release an automated red-teaming (ART) framework designed to measure how safety classifiers affect downstream model behavior in adversarial multi-turn interactions. Using synthetic dialogues generated via a controlled two-agent setup, we train MindGuard, a family of lightweight safety classifiers (with 4B and 8B parameters). Our classifiers reduce false positives at high-recall operating points and, when paired with clinician language models, help achieve lower attack success and harmful engagement rates compared to general-purpose safeguards. We release all models, human evaluation data, and the ART framework.

Large language models (LLMs) are increasingly deployed as interactive agents, yet their capacity for social and strategic reasoning over extended interaction remains poorly understood. Existing evaluations rely on static vignettes or single-game benchmarks that cannot capture the sustained, multi-faceted reasoning that real-world multi-agent settings demand. We introduce Mindgames, a multi-game arena and evaluation platform for LLM agents that operationalizes complementary reasoning demands relevant to ``theory of mind'': belief attribution under hidden information, opponent modeling through repeated strategic interaction, cooperative inference under knowledge asymmetries, and sustained deception in social deduction. Built on TextArena, Mindgames provides a unified interaction interface, TrueSkill-based rating, and full trajectory logging across four game environments. We instantiate Mindgames through a 2025 competition cycle hosted at a major AI conference, which assessed 944 submitted agents from 76 teams across four games: Colonel Blotto, Iterated Prisoner's Dilemma, Codenames, and Secret Mafia. Our analysis surfaces both agent-level and evaluation-level limitations: brittle rule adherence remains a major bottleneck, top-performing systems repeatedly rely on explicit structural scaffolding, and leaderboard validity differs sharply across environments. In particular, failure-heavy environments can reward robustness to opponent errors as much as strategic ability, with Secret Mafia exhibiting a pronounced error-survival confound in this cycle. We release a dataset of 29,571 multi-agent games (comprising 94,132 player trajectories and 243M tokens) with turn-level observations, actions, and rewards, together with MG-Ref, a deterministic offline tournament protocol that scores new agents against a frozen reference pool of top-ranked, low-error Stage~II submissions under the same error-attribution lens used in this analysis. By releasing the benchmark interface, dataset, MG-Ref, and starter kit, Mindgames provides a reproducible resource for studying progress in multi-agent social and strategic reasoning.

Visual decoding from brain signals is a key challenge at the intersection of computer vision and neuroscience, requiring methods that bridge neural representations and computational models of vision. A field-wide goal is to achieve accurate, generalizable decoding from non-invasive, temporally resolved signals, including electroencephalography (EEG). A major obstacle towards this goal is the low signal-to-noise ratio of EEG and the substantial inter-subject variability, which render direct end-to-end EEG–image supervision weak and unstable. To address this, we introduce a tri-modal contrastive framework for EEG-based visual decoding that aligns EEG, visual, and textual representations within a unified latent space. Our approach follows a two-stage design. First, we pre-train an EEG encoder via masked reconstruction on unlabeled trials, learning spatio-temporal regularities that transfer robustly downstream. Second, we jointly align EEG, image, and LLM-generated textual descriptions through contrastive learning, where text supervision acts as a semantic regularizer that injects linguistic structure into the shared space without overwhelming the primary EEG–image signal. The encoder integrates subject-specific adaptation, graph-attention over channels, and temporal-spatial convolutional embeddings. On the Things-EEG2 200-way zero-shot benchmark, our framework achieves 54.1\% Top-1 and 83.4\% Top-5 accuracy, substantially exceeding the strongest prior baseline (32.4\% / 64.0\%), with paired Wilcoxon tests confirming significance (p < 0.01) over all in-subject baselines. We validate generalization on Things-MEG. Analysis reveals that compact embedding geometries (CN-CLIP) outperform much larger backbones, and that decoding aligns with established neurophysiology of visual processing. This work is a critical step towards robust, semantically-grounded visual decoding from non-invasive temporal neural signals.

Representation alignment has emerged as an effective approach to improve Multimodal Large Language Models (MLLMs) by regularizing their internal representations toward those of an external vision encoder. However, existing methods typically align a fixed layer of the language backbone, overlooking the fine-grained structure of Transformer models. In this work, we propose **He**ad-Wise **R**epresentation **A**lignment (**HeRA**), a method that enforces cross-modal alignment at the level of individual attention heads. Our approach is grounded in the Platonic Representation Hypothesis, focusing on preserving the *topological structure* of representations (*i.e.*, their local neighborhood relationships) across modalities. Following the Mutual K-Nearest Neighbor (MKNN) alignment metric, we introduce a contrastive objective that acts as a differentiable proxy for matching local structures. HeRA applies this objective during multimodal training to specific attention heads in the LLM, selected by their alignment score according to the MKNN metric. Counterintuitively, we find that aligning the *least* aligned heads yields the largest gains. Extensive evaluations across multiple MLLMs and 18 benchmarks demonstrate that HeRA consistently improves performance on challenging vision-centric tasks and serves as an effective regularizer against visual hallucinations by naturally curbing the over-reliance on linguistic priors. Our code will be publicly released.

In cooperative multi-agent reinforcement learning under partial observability, agents typically differ in both their local observations and in how much value those observations let them recover. Most existing methods treat agents symmetrically when shaping credit and exploration, ignoring this asymmetric information disadvantage. We introduce the per-agent value gap, a state- and agent-specific quantity that measures the value loss attributable to local information limits, and show that it serves as a single, measurable signal that unifies credit assignment and exploration: down-weighting policy updates where an agent is at a strong disadvantage, and rewarding behaviors that reduce the disadvantage along the trajectory. We instantiate this idea as \method{}, which estimates the gap online with a per-agent local value head and uses it to drive both advantage shaping and an information-seeking intrinsic reward. We further show that the resulting shaping reduces the variance of the policy-gradient estimate at high-disadvantage states under a mild correlation assumption, and that the intrinsic bonus vanishes in expectation under any stationary optimal policy, so the augmented objective remains consistent with the original cooperative objective. On sixteen tasks across five cooperative benchmarks, \method{} ranks first or statistically tied for first on fifteen of sixteen tasks (strictly first on twelve), with the largest gains on settings combining sparse rewards and asymmetric observability (e.g., a $24\%$ mean relative gain over the strongest baseline averaged across the three SMAC v2 tasks).

Audio descriptions (AD) are essential for making visual media accessible to blind and low-vision (BLV) users. While Multimodal Large Language Models (MLLMs) offer a scalable solution for generating audio descriptions on-demand, their performance on ``inline'' audio descriptions -- which must fit within existing silences in a video -- remains under-explored, particularly for diverse user-generated content (UGC). We present a comprehensive investigation into MLLM-generated inline audio descriptions. First, we introduce a professionally annotated dataset of 36 hours of videos, each averaging $\sim$5 mins. and total 7k+ descriptions, providing a high-quality benchmark for this task. Second, we propose an evaluation framework based on audio description expert guidelines that provides fine-grained actionable metrics for model improvement. Crucially, our framework evaluates the end-to-end task: it renders the generated scripts into audio tracks to assess their fit within the original video's natural silences, minimizing disruptive overlaps. Our analysis reveals that while frontier MLLMs accurately describe visual events, significant gaps persist in audio-visual timing, quality, and narrative flow. Finally, we validate the framework and conduct studies with BLV participants and sighted raters, finding that while users narrowly prefer human-authored audio descriptions, they still find MLLM-generated audio descriptions highly beneficial for comprehension. We provide guidance on the current capabilities of MLLMs for audio descriptions and release our dataset and evaluation suite to the community.

Do behavioral styles in large language models correspond to reusable directions in residual space, or do they mainly appear after averaging many noisy contrastive examples? We introduce MinSteer, a two-stage cached-continuation method for estimating a contrastive residual direction (CRD) within one decoding trajectory. Stage-X generates a continuation in one style; a style-flip suffix is then processed using Stage-X's final KV cache; Stage-Y continues from this inherited trajectory. The CRD is computed from generated-token residuals only, excluding prompt and suffix tokens from the average. Across sentiment transfer, politeness control, and a ParaDetox-based bidirectional diagnostic, MinSteer extracts effective steering directions from very small contrasts. On TweetEval sentiment transfer, one MinSteer contrast outperforms a CAA baseline averaged over 50 independent pairs. Japanese keigo CRDs transfer to English and Chinese outputs and outperform trajectory-free TwoSeq controls. In ParaDetox, MinSteer produces bidirectional Detoxify movement under our rewrite setup, while exposing semantic-preservation trade-offs. The directions transfer across tested scenarios and languages and show clear depth dependence. Together, these results show that cached self-rewrite trajectories provide a cleaner signal for residual-space behavior directions than independent-pair averaging.

Large language models increasingly rely on long Chain-of-Thought (CoT) to solve complex reasoning tasks, but the resulting CoT often contains repeated checks, abandoned branches, and unused planning that increase inference cost without supporting the final answer. Existing efficient reasoning methods address this through length budgets, token-level importance, confidence signals, or a stronger teacher model, treating CoT compression mainly as a text-shortening problem rather than as a question of which steps the answer actually depends on. We propose MinPath, a training framework that treats CoT compression as a graph problem. The raw CoT is represented as a Typed Reasoning Graph, where the Minimal Dependency Path supporting the target conclusion is extracted, and the resulting MinPath is used for fine-tuning. Across four reasoning models on GSM8K, MATH-500, and AIME24, MinPath reduces generated tokens by 33.9\% on average with nearly unchanged precision (accuracy even improved in 41.7\% of all experiments) and outperforms five compression baselines in every evaluated setting.

We show that the MinMax algebra provides a form of recurrence that is expressively powerful, efficiently implementable, and most importantly it is not affected by vanishing or exploding gradient. We call MinMax Recurrent Neural Cascades (RNCs) the models obtained by cascading several layers of neurons that employ such recurrence. We show that MinMax RNCs enjoy many favourable theoretical properties. First, their formal expressivity includes all regular languages, arguably the maximal expressivity for a finite-memory system. Second, they can be evaluated in parallel with a runtime that is logarithmic in the input length given enough processors—and they can also be evaluated sequentially. Third, their state and activations are bounded uniformly for all input lengths. Fourth, at almost all points, their loss gradient exists and it is bounded. Fifth, they do not exhibit a vanishing state gradient: the gradient of a state w.r.t. a past state can have constant value one regardless of the time distance between the two states. Finally, we find empirical evidence that the favourable theoretical properties of MinMax RNCs are matched by their practical capabilities: they are able to perfectly solve a number of synthetic tasks, showing superior performance compared to the considered state-of-the-art recurrent neural networks; also, we train a MinMax RNC of 127M parameters on next-token prediction, and the obtained model shows competitive performance for its size, providing evidence of the potential of MinMax RNCs on real-world tasks.

Automated discovery of governing equations remains challenging when models rely only on numerical tokens and ignore the phase-space and trajectory evidence that is informative in low-dimensional dynamics. We propose Visual Induction for Physics-based Equation Reasoning (VIPER-R1), a multimodal framework for visual formula discovery from phase portraits, trajectory plots, and aligned measurements. VIPER-R1 uses a supervised Motion Structure Induction (MSI) curriculum for joint rationale-and-equation generation, and is then refined by \textbf{Reward-Guided Symbolic Calibration (RGSC)}, a GRPO-based stage optimized with a structure-aware reward that emphasizes topological correctness over coefficient matching. At inference time, we optionally apply Symbolic Residual Realignment (SR$^2$), an external symbolic regression module that models only the residual numerical mismatch of the base hypothesis. We release PhysSymbol, a 10{,}000-instance two-part multimodal corpus; the quantitative evaluation in this paper uses its Part 1 mechanics split, which provides paired visualizations, trajectory data, and expert C-CoT traces. On this benchmark, VIPER-R1-7B achieves the best structural score ($S_{\mathrm{struct}}=0.812$), is effectively tied with VIPER-R1-3B on exact-match accuracy ($S_{\mathrm{acc}}=0.487$ vs.\ $0.488$), and, after optional SR$^2$ refinement, reaches a Post-SR$^2$ MSE of $0.032$, nearly $3\times$ lower than the best VLM baseline. These results show that VIPER-R1 improves symbolic structure induction in focused low-dimensional dynamics by combining paired visual evidence with structure-aware calibration to produce stronger first-order hypotheses for downstream refinement.

Transformer-based video generative models have become increasingly realistic, but long-horizon consistency remains challenging to achieve because even a few dozen frames create impractically long sequence lengths. We show that this issue can be mitigated by generating video using coarse-to-fine rollout within a multi-scale token space. Our approach is simple: first, we pre-train an adaptive autoencoder that compresses each frame into a hierarchy of tokens, with levels ranging from the typical latent resolution to only a handful of tokens per frame. This yields a hierarchy in which the coarsest levels capture the most consequential information—such as scene layout and semantics—while finer levels add high-frequency appearance and texture. Then, we train a video diffusion model to generate these tokens using coarse-to-fine rollout. By carefully controlling the level of detail at which frames are generated and used as context during each rollout step, we are able to preserve long-range consistency in geometry and object permanence while expending less compute on long-term consistency of less perceptually relevant details. We validate this approach using a custom long-horizon Minecraft dataset, where it produces substantially more consistent rollouts compared to strong baselines. We hope this perspective opens new opportunities for long-term consistent video generation.

Multimodal large language models (MLLMs) show remarkable potential for scientific reasoning, yet their performance in specialized domains such as microscopy remains limited by the scarcity of domain-specific training data and the difficulty of encoding fine-grained expert knowledge into model parameters. To bridge the gap, we introduce MicroWorld, a framework that constructs a multimodal attributed property graph (MAPG) from large-scale scientific image-caption corpora and leverages it to augment MLLM reasoning at inference time without any domain-specific fine-tuning. MicroWorld extracts biomedical entities and relations via scispaCy or LLM-based triplet mining, aligns images and entities in a shared embedding space using Qwen3-VL-Embedding, and assembles a knowledge graph comprising approximately 111K nodes and 346K typed edges spanning eight relation categories. At inference time, a graph-augmented retrieval pipeline matches query entities to the MAPG and injects structured knowledge context into the MLLM prompt. On the MicroVQA benchmark, MicroWorld improves the reasoning performance of Qwen3-VL-8B-Instruct by 37.5%, outperforming GPT-5 by 13.0% to achieve a new state-of-the-art. Furthermore, it yields a 6.0% performance gain on the MicroBench benchmark. Extensive experiments demonstrate the enhanced generalization capability introduced by MicroWorld. A qualitative case study further reveals both the mechanisms through which structured knowledge improves reasoning and the failure modes that point to promising future directions. Code and data are available at anonymous GitHub.

Diffusion-weighted MRI (dMRI) requires densely sampling $q$-space across many directions and shells, leading to long scan times that limit clinical use. Angular super-resolution (ASR) recovers a dense signal from a sparse acquisition, enabling substantially shorter scans. We propose a physics-guided flow matching framework for ASR built on a basic fact about dMRI: under the cumulant expansion, the diffusion signal naturally splits into a known linear function of the $q$-vectors plus a residual carrying higher-order angular detail. We exploit this twice. First, we initialize the flow from a closed-form dense DTI estimate, recasting ASR as residual flow matching from a physically meaningful starting point rather than uninformative noise. Second, we impose a microstructure-consistency loss, rooted in the physical principle that any two diffusion-weighted signals from the same voxel must be explained by a single underlying set of microstructural parameters. Together, these designs anchor predictions wherever the acquisition is informative while leaving the learned prior to model crossing-fiber and higher-order angular detail. On the Human Connectome Project (HCP-YA) and UK Biobank, our method matches or surpasses analytical baselines and recent deep learning methods on signal reconstruction, kurtosis scalar maps, and fiber-orientation reconstruction under aggressive subsampling.

Modeling traffic dynamics is a critical challenge for urban computing, with applications from real-time traffic management to infrastructure planning. However, progress in this area is fundamentally constrained by a lack of large-scale public datasets that capture the subtle properties of real city road networks. Existing benchmarks are often limited by their small scale, reliance on sparse highway traffic sensors, absence of true road connectivity information, and lack of information about road properties. To address this issue, we introduce datasets representing fine-grained road networks of two major cities, which are unique in their scale (up to 100,000 road segments), use of real road connectivity, presence of time series measurements for both traffic speed and volume at a 5-minute resolution, and inclusion of rich static road attributes. These datasets enable in-depth analysis of spatiotemporal traffic patterns and can serve as benchmarks for various ML applications. As a practical demonstration of the utility of our datasets and the challenges they present, we use them for the task of traffic forecasting. The size of the real-world road networks in our datasets reveals significant scalability issues in current traffic forecasting models. To address them, we propose a simple and efficient baseline that not only scales to large road graphs but also achieves forecasting performance competitive with other established spatiotemporal models. We hope that the proposed datasets will serve as a foundational resource for a broad range of research in traffic modeling, urban computing, and smart city development.

Sampling from score-based diffusion models incurs bias due to both time discretisation and the approximation of the score function. A common strategy for reducing this bias is to apply corrector steps based on the unadjusted Langevin algorithm (ULA) at each noise level within a predictor-corrector framework. However, ULA is itself a \emph{biased} sampler, as it discretises a continuous diffusion process. In this work, we consider \emph{adjusted} Langevin correctors that employ Metropolis--Hastings (MH) or Barker's accept-reject steps to correct for this bias. Since the target density ratio typically required by MH-based algorithms is unavailable, we propose methods that instead utilise the score function to compute the correct acceptance probability. We introduce the first exact method for adjusting Langevin corrections in diffusion models, based on a two-coin Bernoulli factory algorithm. We also propose an efficient approximation based on Simpson's rule that achieves accuracy of order $5/2$ in the step size at near-zero marginal cost. We demonstrate that these procedures improve sample quality on both synthetic and image datasets, yielding consistent gains in Fr\'echet Inception Distance on the latter.

We propose AdaDS, a generalizable framework for prompted metric depth estimation, which estimates high-resolution metric depth from images and arbitrarily degraded low-resolution depth prompts. This setting is commonly studied as depth super-resolution, where existing methods typically regress depth values directly and often exhibit artifacts under severe or unknown depth degradation. In contrast, AdaDS exploits the contraction property of Gaussian smoothing: as noise accumulates in the forward diffusion process, the distributional discrepancy between degraded depth prompts and their high-quality counterparts progressively diminishes, eventually approaching an isotropic Gaussian prior. Leveraging this property, AdaDS estimates refinement uncertainty to adaptively select a starting timestep in the reverse diffusion trajectory, and subsequently injects tailored noise to place the intermediate sample in a high-probability region of the target posterior distribution. This strategy enables the generative prior of a pre-trained diffusion model to dominate the estimation process even when upstream prompt refinements are imperfect. Extensive experiments on real-world and synthetic benchmarks demonstrate AdaDS's superior zero-shot generalization and robustness to diverse degradation patterns compared with state-of-the-art methods.

Large language models are increasingly deployed in safety-critical and user-facing applications, where their ability to resist harmful instructions is essential. Although post-training aims to make models robust against many jailbreak strategies, recent evidence shows that stylistic reformulations, such as poetic transformation, can still bypass safety mechanisms with alarming effectiveness. This raises a central question: why do literary jailbreaks succeed? In this work, we investigate whether their effectiveness depends on specific poetic devices, on a failure to recognize literary formatting, or on deeper changes in how models process stylistically irregular prompts. We address this problem through an interpretability analysis of attention patterns. Our analysis proceeds in three steps: we perform input-level ablation studies to assess the contribution of individual and combinations of rhetorical devices; we construct a novel interpretable vector representation of attention maps; we cluster these representations and train linear probes to predict both safety outcomes and literary format. Our results show that models distinguish poetic from prose formats with high accuracy, yet struggle to predict jailbreak success within each format. Clustering further reveals clear separation by literary format, but not by safety label. These findings indicate that jailbreak success is not caused by a failure to recognize poetic formatting; rather, poetic prompts induce distinct processing patterns that remain largely independent of harmful-content detection. Overall, literary jailbreaks appear to misalign large language models not through any single poetic device, but through accumulated stylistic and structural irregularities that alter prompt processing and avoid lexical triggers considered during post-training. This suggests that robustness requires safety mechanisms that account for style-induced shifts in model behavior. We use Qwen3-14B as a representative open-weight case study for all reported experiments.

We introduce MetaLoop, a seven-task benchmark that evaluates the full metacognitive loop in large language models: monitoring one's own uncertainty, translating that signal into action (abstaining, switching strategies, correcting errors), and updating from feedback. Existing evaluations test these components in isolation; MetaLoop produces integrated per-model profiles across all three. Evaluating 12 frontier models alongside 460 human participants, we find that (1) accuracy does not predict metacognitive ability (high-accuracy models routinely approve their own errors); (2) models universally produce calibration signals but fail to act on them; (3) forced self-explanation distinguishes genuine self-monitoring from brittle pattern-matching; and (4) humans and LLMs both show monitoring–control gaps but with different failure modes (humans are loss-averse, models overconfident), with clarification detection as the one clear human advantage. We release the benchmark, scoring code, model outputs, and human data.

Existing locate-then-edit Knowledge Editing (KE) methods typically decompose editing into two stages: upstream target representation optimization and downstream constrained parameter optimization. The optimization across the two stages is disconnected: upstream applies uniform regularization without observing downstream realization of the planned residual, hindering a refined accuracy–editability trade-off. Since this realization is request-specific and depends on downstream constraints, uniform regularization can over-shrink high-association requests, causing insufficient editing, while it can under-regularize low-association requests, producing over-large planned residuals that reduce downstream editability. To bridge this disconnect, we propose $\textbf{MetaKE}$ ($\textbf{Meta}$-learning for $\textbf{K}$nowledge $\textbf{E}$diting), a new framework that unifies upstream and downstream stages into a bi-level optimization problem. The inner level optimizes parameter updates for the target representation, while the outer level optimizes representation using feedback from downstream constraints, achieving a better semantic accuracy-editability trade-off. To avoid costly multi-layer backpropagation, we introduce a Structural Gradient Proxy to approximate and propagate this feedback. Extensive experiments show that MetaKE outperforms strong baselines, offering a new perspective on KE. Code at: https://anonymous.4open.science/r/MetaKE-C160.