NeurIPS 2026 论文列表

Test-time sampling can improve Vision-Language-Action (VLA) policies, but only if the model can efficiently select a good action from multiple candidates. Prior work often relies on separately trained external verifiers, adding extra models and inference overhead. We propose \textbf{Token Probability Bucketing (TokenPB)}, a self-verification approach that discretizes continuous action chunks into bucket tokens and co-trains a single VLA checkpoint with a joint objective: a flow-matching loss for action generation and an auxiliary bucket loss $\mathcal{L}_{\text{bucket}}$ that teaches the VLM backbone to model observation-conditioned distributions over bucket-token sequences. At inference, TokenPB samples $M$ candidates, scores each via teacher-forced evaluation under the learned distribution, and selects the best---enabling efficient batched scoring with observation-prefix KV-cache reuse. We provide theory connecting the training objective to a monotonic ranking property and bounds on expected selection regret that separate candidate-set effects from calibration. Across multiple backbones and benchmarks, TokenPB yields consistent success-rate improvements over single-sample inference from the same checkpoint (+2.7 in simulation, +17.1 in real-world, +16.4 under out-of-distribution shifts), and further improves over external-verifier baselines under the same sampling budget.

Text-to-image (T2I) diffusion models are increasingly distilled into few-step variants and being deployed to enable fast inference. However, their ability to generate harmful or undesired content poses significant safety risks. Data-driven unlearning methods suppress targeted generations by fine-tuning model weights using specialized unlearning objectives. Crucially, these objectives implicitly rely on multi-step denoising dynamics, an assumption that breaks down for few-step distilled (FSD) models, resulting in ineffective forgetting. Furthermore, performing unlearning on the non-distilled base model and subsequently re-distilling it to obtain an unlearned FSD model incurs substantial computational and time overhead, making it impractical in many settings. Hence, we address this limitation with a preference-driven unlearning framework that revisits Direct Preference Optimization (DPO) for diffusion models. We show that standard DPO and its unlearning derivatives, formulated around noise-prediction error, transfer poorly to FSD models due to their altered generation dynamics. To overcome this, we introduce a modified preference optimization formulation explicitly aligned with the few-step generation properties, $\textit{enabling direct concept removal in FSD models}$ while preserving few-step efficiency and maintaining strong retention of desirable (non-targeted) capabilities. We evaluate our framework primarily on identity and NSFW (nudity) removal tasks and also extend our method to object-level unlearning. Extensive experiments demonstrate consistent and effective forgetting, and strong retention performance, establishing our method as a practical and principled solution for unlearning in FSD models.

Denoising score matching (DSM) is the cornerstone of scalable diffusion model training. However, DSM becomes inapplicable on curved manifolds: the curved geometry makes the forward diffusion process nonlinear and the transition kernel is no longer tractable. Existing approaches therefore revert to implicit score matching (ISM) with expensive Brownian simulation, approximate the intractable heat kernel, or bypass the transition kernel altogether via flow matching with the logarithm map, each incurring significant computational cost or approximation error on general manifolds. We propose \emph{splitting diffusion}, a new framework that restores DSM-like training on general Riemannian manifolds by replacing the intractable Brownian forward process with a splitting scheme that alternates two steps: (1) an Ornstein-Uhlenbeck velocity update in the tangent space, whose transition kernel has a closed-form solution; (2) a deterministic geodesic transport step. This splitting yields a fully tractable transition kernel locally, enabling DSM-style training while requiring only the exponential map as a geometric oracle. The effectiveness of splitting diffusion is demonstrated on complicated manifolds and high-dimensional data.

While existing time series foundation models primarily rely on large-scale unimodal pretraining, they lack complementary modalities to enhance time series understanding. Building multimodal foundation models is a natural next step, but it faces key challenges: 1) lack of a unified multimodal pretraining paradigm and large-scale multimodal corpora for time series analysis; 2) how to effectively integrate heterogeneous modalities and enhance model generalization. To address these challenges, we take an early step toward multimodal foundation models for time series analysis. We first propose a multimodal pretraining paradigm that leverages time series with endogenous modalities (derived images and text) and exogenous knowledge (news), providing a comprehensive multi-view perspective for time series analysis. To support this, we develop an automated data construction pipeline to curate MM-TS, the first large-scale multimodal time series dataset spanning six domains, with up to one billion points. Then we propose HORAI, a frequency-enhanced multimodal foundation model. It integrates two core components: the Frequency-enhanced Cross-Modality Encoder and the Time-Frequency Decoder, designed to effectively fuse multimodal features and enhance model generalization across modalities and domains. After pretraining on MM-TS, HORAI achieves state-of-the-art zero-shot performance on time series forecasting and anomaly detection tasks, demonstrating strong generalization.

Diffusion models promise iterative refinement, yet masked diffusion language models (MDMs) suffer from two shortcomings: (1) inability to revise tokens, and (2) sparse training signal. The root cause is structural: MDMs control information flow only through input masking, so unmasked positions cannot be supervised, and self-correction is never learned. We propose the Leave-One-Out Transformer (LOOT), which addresses both issues. With LOOT, MDMs predict the distribution over potential replacements conditioned on the rest of the sequence at each position, in a single forward pass. Supervision can then be applied at masked and unmasked positions alike, yielding a lower-variance training objective with a dense loss signal at every token. Our method empowers MDMs to learn self-correction during training and supports continuous refinement at inference time via Gibbs sampling. LOOT achieves a new best validation perplexity on LM1B in a parameter-matched setting. Finetuned from a MDM checkpoint on OpenWebText, LOOT establishes a generative quality-vs-diversity frontier that surpasses prior remasking methods across NFE budgets.

When large language models (LLMs) are used for creative tasks where the evaluation is largely subjective, two crucial activities are ideation (in which a list of candidate ideas is produced) and pruning (in which the initial candidates are prioritized through a process of comparison). In this framework, the subjective results are influenced by inherent biases in the LLM's generation and evaluation, and it is therefore important to understand how these subjective biases operate. We define and explore a set of stylized tasks that allow us to highlight the effects of these subjective biases in LLM evaluation. In the process, we find evidence for several key underlying principles. First, there are multiple subjective LLM biases at play, and to understand the overall evaluation we must analyze the relative strengths of these biases and their interactions. Second, different prompts -- even when they seem superficially similar -- can have the effect of making different subjective LLM biases more or less salient. And third, there are complex interactions between the LLM's initial idea-generation phase and the subsequent decisions that go into prioritizing and pruning these ideas.

We study methods for simultaneous analysis of many noisy and biased estimates, each paired with an even noisier estimate of its own bias. The analyst's goal is to construct short calibrated intervals for each parameter. The standard debiasing approach, which subtracts the bias estimate from each biased estimate, inflates variance and yields long intervals. In this paper, we propose an empirical Bayes rebiasing strategy that starts from the fully debiased estimates and learns from data how much bias to reintroduce by estimating the unknown bias distribution. We provide convergence rates for the coverage of our intervals when the bias distribution is estimated using nonparametric maximum likelihood. Furthermore, we demonstrate substantial precision gains in prediction-powered inference, including pairwise LLM win-rate evaluations, as well as for inference of direct genetic effects in family-based GWAS.

Empathetic spoken dialogue is a sophisticated cognitive process that requires not only recognizing emotions but also inferring a user’s latent mental states to provide appropriate support. However, current SpeechLLMs often treat empathy as a direct input-to-response mapping, leading to "superficially warm" but emotionally hollow interactions. In addition, since empathy relies on a multi-stage process with strong inter-step dependency, errors at any intermediate step can cascade through subsequent steps and lead to inappropriate responses, while existing training paradigms lack mechanisms to precisely localize and improve such errors. In this work, we propose EmpathyChat, a unified framework that reformulates empathetic spoken dialogue as a structured cognitive reasoning process integrating perception, mental-state reasoning, and response generation. To support this paradigm, we first construct EmpathySet-400K, an acoustically rich dataset for multi-stage empathetic supervision. During the SFT stage, we strengthen acoustic grounding through proposed Acoustic-Anchored Attention (AAA). During the RL stage, we further introduce a novel stage-aware optimization objective with Step-Decomposed Credit Assignment (SDCA) to localize reasoning errors and mitigate cascaded error propagation. In addition, we introduce EmpathyEval, an expert-annotated benchmark for multi-dimensional empathy evaluation. Extensive experiments demonstrate that EmpathyChat achieves state-of-the-art performance in perception, reasoning, and response alignment. All resources, including model, code, dataset, and benchmark, will be publicly released.

Humans have a remarkable ability to judge causal relationships from a limited number of unreliable observations. Past work on causal cognition has largely focused on normative accounts of human behavior, leaving unknown how biologically plausible neural systems could learn causal relationships from observations and update their representations of causal structure with additional evidence. Here, we leverage task-optimized recurrent neural networks to discover candidate implementation-level neural mechanisms of causal judgment. We propose a novel cognitive task in which a subject observes stochastic samples from an unknown causal structure (e.g. among variables A, B, and C with unknown causal relationships), and must judge whether a specific causal relationship is present given a query (e.g. "does A cause C?"). We found that, after training, recurrent neural networks perform the task with high accuracy, adopt strategies that incorporate the behavior of non-queried variables to form their judgments, and, despite being trained only on pairwise queries ("does A cause B?", "does C cause A?", etc), form implicit beliefs about the complete graphical structure underlying the observations. Lastly, we use dynamical systems analysis to identify a set of low-level neural mechanisms that implement causal judgment and representation of causal graphical structure. Together, these findings lay the groundwork for a "bottom-up" approach to causal cognition, providing a potential basis for subsequent experimental study in the brain.

Understanding how linguistic structure emerges in language models is central to interpreting what these systems learn from data and how much supervision they truly require. In particular, semantic role understanding (``who did what to whom’’) is a core component of meaning representation, yet it remains unclear whether it arises from pre-training alone or depends on task-specific fine-tuning. In this paper, we study whether semantic role understanding emerges during language model pre-training or requires task-specific fine-tuning. We freeze decoder-only transformers and train linear probes to extract semantic roles, using performance to infer whether role information is already encoded in pre-training or learned during adaptation. Across model scales, we find that frozen representations contain substantial semantic role information, with performance improving but not fully matching fine-tuned models. This indicates partial but incomplete emergence from pre-training alone. We show that semantic role structure emerges from language modeling objectives, but its internal implementation shifts toward more distributed representations as model scale increases.

Fine-tuning LLMs on narrow harmful datasets can induce Emergent Misalignment (EM), where models exhibit misaligned behavior far beyond the fine-tuning distribution. We argue that emergent misalignment can be better understood as a data-mediated transfer phenomenon: harmful fine-tuning examples do not induce uniform behavioral spillover, but interact with the structural properties of the dataset and the difficulty of the tasks relative to the model. Across our experiments, we find that misalignment appears more readily when fine-tuning and evaluation prompts share similar underlying functional structure, when prompts leave more room for coherent harmful completions, and when the target behavior has been more reliably learned by the model. The training pipeline itself also matters: pretraining composition shapes later misalignment. We further study Subliminal Learning (SL), where misalignment is transmitted by fine-tuning on seemingly benign data generated by a harmful teacher. Moving beyond the standard SFT setting, we for the first time compare this transfer under off-policy and on-policy distillation as well, allowing us to separate the roles of the teacher guidance and the training data distribution in transmitting misalignment. Together, these results argue for a data-centric view: Emergent/subliminal misalignment should not be treated as a simple consequence of isolated harmful fine-tuning examples, but as the result of interactions between fine-tuning data structure, pretraining distributions, and training channels.

Kolmogorov–Arnold Networks (KANs) expand each input coordinate through a fixed basis family, yet the optimization role of this basis lifting remains poorly understood. We identify basis lifting as a structural source of ill-conditioning in KANs and provide a mechanism-level analysis linking lifted feature geometry to convergence, accuracy, and seed sensitivity. Our analysis studies the evolution of lifted empirical Gram matrices under stochastic gradient descent (SGD) and shows that fixed basis lifting induces and preserves near-rank deficiency through eigenvalue and determinant drift. It further concentrates feature energy into a smaller number of directions, amplifying dominant eigenvalues and worsening condition numbers. These results show that basis selection is not merely an expressivity choice, but also an optimization and conditioning choice. Motivated by this mechanism, we introduce T-KAN, a learnable basis-coordinate transformation that aligns basis directions before the learned linear map. It reshapes lifted Gram geometries while preserving the expressivity and introducing negligible parameter overhead. In controlled small-to-moderate KAN regimes, T-KAN often improves conditioning and optimization with B-spline, Chebyshev, and RBF KANs.

Classifier-free guidance (CFG) is the de facto standard for conditional sampling in diffusion models, yet it often reduces sample diversity. Using tools from statistical physics, we analyze the emergence of generative distortions induced by CFG, namely the mismatch between the CFG sampling distribution and the true conditional distribution. We study this phenomenon in analytically tractable settings with exact score functions, characterizing its dependence on data dimensionality and the number of classes. For high-dimensional Gaussian mixtures, we use dynamic mean-field theory to show that distortions arise when the number of classes scales exponentially with the data dimension, whereas they vanish in the sub-exponential regime due to a dynamical phase transition. We further prove that, in the infinite-class limit, distortions remain unavoidable regardless of dimensionality because of the increasing density of classes. Finally, we show that standard CFG schedules cannot prevent variance shrinkage, and we propose a theoretically grounded guidance schedule incorporating a negative-guidance window that improves both class separability and sample diversity in real-world latent diffusion models.

Recent advances in 3D object generation have enabled the creation of high-fidelity 3D assets from 2D images. This *2D-to-3D* capability offers a promising path for alleviating 3D data scarcity, which remains a key bottleneck for scalable simulation in robotics and AR/VR applications. However, existing 3D generation methods primarily emphasize visual plausibility and geometric fidelity, with limited consideration of object-level interaction and structural understanding that are critical for downstream embodied AI tasks. To bridge this gap, we introduce **EmbodiedObject**, a unified framework for embodied 3D object understanding built upon latent representations produced by modern *2D-to-3D* generative models such as SAM3D. We first systematically analyze the part-level locality of these latent representations against existing unified 3D point encoders and show that they naturally preserve rich structural semantics suitable for object understanding. Motivated by this observation, we then design a unified DiT-style decoder that operates directly on the latent representation and supports multiple object-level tasks, including open-vocabulary 3D affordance prediction, 3D part segmentation, and 3D articulation estimation. Given a real-world 2D image containing an object of interest, **EmbodiedObject** extends *2D-to-3D* generation beyond visual reconstruction by jointly reasoning about affordance, articulation, and part semantics. This transforms generated 3D assets into actionable object representations suitable for downstream embodied AI applications. Our source-code will be open-sourced.

Modern AI is opening the door to collective decision-making in which participants express their views as free-form text rather than voting on a fixed set of candidates. A natural idea is to embed these opinions in a vector space so that the substantial literature on facility location problems and fair clustering can be brought to bear. But standard text embeddings measure semantic similarity, whereas distances in facility location problems and fair clustering require what we call preferential similarity: a participant's agreement with a piece of text should be inversely related to their distance from it. Off-the-shelf embeddings inherit a coarse preference signal through a correlation between semantic and preferential similarity, but fail to capture preferences when the correlation breaks. We formalize this as an invariance problem: text embedding models encode both a preference-relevant signal (stance and values) and semantic nuisance (style and wording), and the two are observationally correlated, so a geometry that relies on nuisance can appear preference-correct even when it is not. We show that synthetic training data designed to break this correlation provably shifts the optimal scorer away from nuisance-dominated cosine and significantly improves preference prediction across 11 online deliberation datasets.

AI-enabled Cyber-Physical Systems (CPS) are highly vulnerable to adversarial and anomalous inputs, where small perturbations can induce cascading errors and unsafe control actions. Existing approaches—such as rule-based filtering, training-time regularization, or diffusion-based reconstruction—either operate outside the model or lack mechanisms to incorporate formal security specifications into the prediction process. In this paper, we develop the first work toward embedding security properties directly into AI-enabled CPS, enabling predictive models to enforce system-level constraints during inference rather than relying on external defenses. We introduce a logic-conditioned bi-stage diffusion framework that integrates Signal Temporal Logic (STL) specifications into forecasting. STL serves as a first-class conditioning signal that guides both an input repair stage and an output refinement stage, allowing the model to jointly mitigate adversarial perturbations and enforce desired temporal behaviors to satisfy security-critical properties. We evaluate our approach on a real-world multivariate CPS forecasting task under both physical sensor and cyber attacks. Our method improves robustness and specification compliance, degrades more gracefully as attack strength increases, and generalizes better to unseen attacks. Ablation studies further show that embedding logical security properties yields gains unattainable by reconstruction-based methods alone, highlighting a new direction for integrating formal methods with generative models in secure CPS.

Tabular foundation models achieve strong accuracy on choice prediction tasks, but their predictions often violate the economic logic those tasks require: raising a price can increase predicted demand, implied willingness-to-pay estimates are frequently negative or implausible, and unavailable alternatives receive nonzero probability. We propose a two-stage adapter that takes a foundation model's predicted choice probabilities as a precomputed feature and embeds them inside a multinomial logit's utility. In Stage 1, we fit the multinomial logit's structural coefficients by maximum likelihood with sign constraints; in Stage 2, we freeze those coefficients and fit a small neural correction operating on the foundation model's predictions. We prove that this composition exactly preserves the multinomial logit's marginal rate of substitution, so analytically computable value-of-time becomes a mathematical guarantee rather than an empirical accident. Across three datasets and two foundation models, the adapter gains 6.4 percentage points (pp) of test accuracy on average over the multinomial logit and up to 12.8 pp, maintains 100\% cost monotonicity, and produces values of time within the published transportation-economics range on the transportation datasets. Performance degrades gracefully under foundation-model context restriction, retaining at least 6 pp of accuracy gain even at 10\% of the original foundation-model context.

Large Language Models (LLMs) have demonstrated remarkable capabilities in Information Extraction (IE) tasks; however, their performance in zero-shot Named Entity Recognition (NER) remains suboptimal. We argue that this limitation stems not from intrinsic defects within the models, but rather from the failure of existing instruction paradigms to effectively elicit their latent capabilities. Drawing inspiration from cognitive science, we propose an instruction optimization framework termed Instruction Semantic Elaboration (ISE), which fully elicits models' latent NER capabilities by mimicking the semantic elaboration process—specifically by contrasting related concepts, decomposing constituent elements, and illustrating with concrete examples. We evaluate our framework under a zero-shot setting across four domain-specific datasets. Results demonstrate that ISE effectively and stably unlocks the intrinsic NER capabilities of LLMs, achieving a 10.99\% improvement in F1 score over initial instructions and outperforming strong baselines by 6.89\%. Furthermore, the framework generalizes well across model architectures and parameter scales, while remaining robust to the quality of initial instructions. This study offers a novel perspective on activating NER capabilities in LLMs, facilitating their efficient and stable deployment in real-world scenarios.

Hour-long videos, streaming feeds, and movie-length content are pushing vision-language models (VLMs) past their breaking point—the standard recipe of encoding each frame into visual tokens yields sequences that grow linearly with video length, causing prohibitive memory consumption and inference latency. Existing approaches, such as token reduction and keyframe selection, attempt to shorten the token sequence but leave the $O(n)$ scaling intact, failing to resolve the fundamental bottleneck. We propose \model, \emph{a strikingly simple, training-free frame packing framework that compresses arbitrarily long videos into a fixed-size visual representation at constant memory cost.} \model requires no fine-tuning, no architectural modification, and no auxiliary models—it plugs directly into off-the-shelf VLMs at inference time, yet consistently surpasses more elaborate baselines. Drawing inspiration from FramePack in generation literature, we propose a new content-aware compression strategy that operates under a fixed token budget for visual understanding. Specifically, \model proceeds in two steps: (1) scoring each frame by its semantic similarity to the text prompt and ranking frames in a diversity-driven manner, and (2) packing the selected frames at varying resolutions into a fixed-length context window. This design concentrates the memory budget on frames that are both instruction-relevant and visually diverse. While in theory \model can pack arbitrarily many frames into the fixed budget, we identify an empirical instantiation that generalizes robustly across different settings. Extensive experiments on LongVideoBench, MLVU, and Video-MME show that \model consistently outperforms full-resolution baselines as well as prior token reduction and keyframe selection methods, across diverse model families, parameter scales, and frame budgets. Notably, on Qwen3.5-35B-A3B with a 128-frame budget, \model yields gains of \textbf{+12.9\%, +10.6\%, and +9.7\%} over the official model on the three benchmarks, respectively.