NeurIPS 2025 Papers

Reinforcement Learning (RL) has proven to be an effective post-training strategy for enhancing reasoning in vision–language models (VLMs). Group Relative Policy Optimization (GRPO) is a recent prominent method that encourages models to generate complete reasoning traces before answering, leading to increased token usage and computational cost. Inspired by the human-like thinking process—where people skip reasoning for easy questions but think carefully when needed—we explore how to enable VLMs to first decide *when reasoning is necessary*. To realize this, we propose \ours, a two-stage training strategy: **(i)** a supervised fine-tuning (SFT) stage with a simple yet effective “**thought dropout**” operation, where reasoning traces are randomly replaced with empty thoughts. This introduces a think-or-not format that serves as a cold start for selective reasoning; **(ii)** a GRPO stage that enables the model to freely explore when to think or not, while maximizing task-aware outcome rewards. Experimental results show that \ours can *reduce the completion length by up to **90%** compared to vanilla GRPO, without sacrificing performance or even improving it*. Further evaluations across LLM (GSM8K), VLM (CLEVR, Super-CLEVR, GeoQA), and Agentic (AITZ) tasks—covering a range of reasoning difficulties under both 3B and 7B models—consistently reveal that the \textit{model progressively learns to bypass unnecessary reasoning steps as training advances}. These findings shed light on the path toward human-like reasoning patterns in RL approaches. Our code is available at https://github.com/kokolerk/TON.

Generative video modeling has made significant strides, yet ensuring structural and temporal consistency over long sequences remains a challenge. Current methods predominantly rely on RGB signals, leading to accumulated errors in object structure and motion over extended durations. To address these issues, we introduce WorldWeaver, a robust framework for long video generation that jointly models RGB frames and perceptual conditions within a unified long-horizon modeling scheme. Our training framework offers three key advantages. First, by jointly predicting perceptual conditions and color information from a unified representation, it significantly enhances temporal consistency and motion dynamics. Second, by leveraging depth cues, which we observe to be more resistant to drift than RGB, we construct a memory bank that preserves clearer contextual information, improving quality in long-horizon video generation. Third, we employ segmented noise scheduling for training prediction groups, which further mitigates drift and reduces computational cost. Extensive experiments on both diffusion and rectified flow-based models demonstrate the effectiveness of WorldWeaver in reducing temporal drift and improving the fidelity of generated videos.

Vertical Federated Learning (VFL) enables participants to collaboratively train models on aligned samples while keeping their heterogeneous features private and distributed. Despite their utility, VFL models remain vulnerable to adversarial attacks during inference. Adversarial Training (AT), which generates adversarial examples at each training iteration, stands as the most effective defense for improving model robustness. However, applying AT in VFL settings (VFAL) faces significant computational efficiency challenges, as the distributed training framework necessitates iterative propagations across participants. To this end, we propose **_DecVFAL_** framework, which substantially accelerates **_VFAL_** training through a dual-level ***Dec***oupling mechanism applied during adversarial sample generation. Specifically, we first decouple the bottom modules of clients (directly responsible for adversarial updates) from the remaining networks, enabling efficient _lazy sequential propagations_ that reduce communication frequency through delayed gradients. We further introduce _decoupled parallel backpropagation_ to accelerate delayed gradient computation by eliminating idle waiting through parallel processing across modules. Additionally, we are the first to establish convergence analysis for VFAL, rigorously characterizing how our decoupling mechanism interacts with existing VFL dynamics, and prove that _DecVFAL_ achieves an $\mathcal{O}(1/\sqrt{K})$ convergence rate matching that of standard VFLs. Experimental results show that _DecVFAL_ ensures competitive robustness while significantly achieving about $3\sim10\times$ speed up.

This paper studies the approximation and generalization abilities of score-based neural network generative models (SGMs) in estimating an unknown distribution $P_0$ from $n$ i.i.d. observations in $d$ dimensions. Assuming merely that $P_0$ is $\alpha$-sub-Gaussian, we prove that for any time step $t \in [t_0, n^{\mathcal{O}(1)}]$, where $t_0 > \mathcal{O}(\alpha^2n^{-2/d}\log n)$, there exists a deep ReLU neural network with width $\leq \mathcal{O}(n^{\frac{3}{d}}\log_2n)$ and depth $\leq \mathcal{O}(\log^2n)$ that can approximate the scores with $\tilde{\mathcal{O}}(n^{-1})$ mean square error and achieve a nearly optimal rate of $\tilde{\mathcal{O}}(n^{-1}t_0^{-d/2})$ for score estimation, as measured by the score matching loss. Our framework is universal and can be used to establish convergence rates for SGMs under milder assumptions than previous work. For example, assuming further that the target density function $p_0$ lies in Sobolev or Besov classes, with an appropriately early stopping strategy, we demonstrate that neural network-based SGMs can attain nearly minimax convergence rates up to logarithmic factors. Our analysis removes several crucial assumptions, such as Lipschitz continuity of the score function or a strictly positive lower bound on the target density.

Scaling visual generation models is essential for real-world content creation, yet requires substantial training and computational expenses. Alternatively, test-time scaling has garnered growing attention due to resource efficiency and promising performance. In this work, we present the first general test-time scaling framework for visual auto-regressive (VAR) models, TTS-VAR, modeling the generation process as a path searching problem. Inspired by VAR's hierarchical coarse-to-fine multi-scale generation, our framework integrates two key components: (i) At coarse scales, we observe that generated tokens are hard for evaluation, possibly leading to erroneous acceptance of inferior samples or rejection of superior samples. Noticing that the coarse scales contain sufficient structural information, we propose clustering-based diversity search. It preserves structural variety through semantic feature clustering, enabling later selection on samples with higher potential. (ii) In fine scales, resampling-based potential selection prioritizes promising candidates using potential scores, which are defined as reward functions incorporating multi-scale generation history. To dynamically balance computational efficiency with exploration capacity, we further introduce an adaptive descending batch size schedule throughout the causal generation process. Experiments on the powerful VAR model Infinity2B show a notable 8.7% GenEval score improvement (0.69→0.75). Key insights reveal that early-stage structural features effectively influence final quality, and resampling efficacy varies across generation scales.

Automated model discovery is the process of automatically searching and identifying the most appropriate model for a given dataset over a large combinatorial search space. Existing approaches, however, often face challenges in balancing the capture of fine-grained details with ensuring generalizability beyond training data regimes with a reasonable model complexity. In this paper, we present a multi-modal \& multi-step pipeline for effective automated model discovery. Our approach leverages two vision-language-based modules (VLM), AnalyzerVLM and EvaluatorVLM, for effective model proposal and evaluation in an agentic way. AnalyzerVLM autonomously plans and executes multi-step analyses to propose effective candidate models. EvaluatorVLM assesses the candidate models both quantitatively and perceptually, regarding the fitness for local details and the generalibility for overall trends. Our results demonstrate that our pipeline effectively discovers models that capture fine details and ensure strong generalizability. Additionally, extensive ablation studies show that both multi-modality and multi-step reasoning play crucial roles in discovering favorable models.

Plant traits such as leaf carbon content and leaf mass are essential variables in the study of biodiversity and climate change. However, conventional field sampling cannot feasibly cover trait variation at ecologically meaningful spatial scales. Machine learning represents a valuable solution for plant trait prediction across ecosystems, leveraging hyperspectral data from remote sensing. Nevertheless, trait prediction from hyperspectral data is challenged by label scarcity and substantial domain shifts (\eg across sensors, ecological distributions), requiring robust cross-domain methods. Here, we present GreenHyperSpectra, a pretraining dataset encompassing real-world cross-sensor and cross-ecosystem samples designed to benchmark trait prediction with semi- and self-supervised methods. We adopt an evaluation framework encompassing in-distribution and out-of-distribution scenarios. We successfully leverage GreenHyperSpectra to pretrain label-efficient multi-output regression models that outperform the state-of-the-art supervised baseline. Our empirical analyses demonstrate substantial improvements in learning spectral representations for trait prediction, establishing a comprehensive methodological framework to catalyze research at the intersection of representation learning and plant functional traits assessment. We also share the dataset\footnotemark[1], code and pretrained model objects for this study \href{https://github.com/echerif18/HyspectraSSL}{here}. \footnotetext[1]{GreenHyperSpectra dataset: \href{https://huggingface.co/datasets/Avatarr05/GreenHyperSpectra}{https://huggingface.co/datasets/Avatarr05/GreenHyperSpectra}}

Transformers have demonstrated remarkable capabilities in multi-step reasoning tasks. However, understandings of the underlying mechanisms by which they acquire these abilities through training remain limited, particularly from a theoretical standpoint. This work investigates how transformers learn to solve symbolic multi-step reasoning problems through chain-of-thought processes, focusing on path-finding in trees. We analyze two intertwined tasks: a backward reasoning task, where the model outputs a path from a goal node to the root, and a more complex forward reasoning task, where the model implements two-stage reasoning by first identifying the goal-to-root path and then reversing it to produce the root-to-goal path. Our theoretical analysis, grounded in the dynamics of gradient descent, shows that trained one-layer transformers can provably solve both tasks with generalization guarantees to unseen trees. In particular, our multi-phase training dynamics for forward reasoning elucidate how different attention heads learn to specialize and coordinate autonomously to solve the two subtasks in a single autoregressive path. These results provide a mechanistic explanation of how trained transformers can implement sequential algorithmic procedures. Moreover, they offer insights into the emergence of reasoning abilities, suggesting that when tasks are structured to take intermediate chain-of-thought steps, even shallow multi-head transformers can effectively solve problems that would otherwise require deeper architectures.

Computational epidemiology is crucial in understanding and controlling infectious diseases, as highlighted by large-scale outbreaks such as COVID-19. Given the inherent uncertainty and variability of disease spread, Monte Carlo (MC) simulations are widely used to predict infection peaks, estimate reproduction numbers, and evaluate the impact of non-pharmaceutical interventions (NPIs). While effective, MC-based methods require numerous runs to achieve statistically reliable estimates and variance, which suffer from high computational costs. In this work, we present a unified theoretical framework for analyzing disease spread dynamics on both directed and undirected contact networks, and propose an algorithm, **RAPID**, that significantly improves computational efficiency. Our contributions are threefold. First, we derive an asymptotic variance lower bound for MC estimates and identify the key factors influencing estimation variance. Second, we provide a theoretical analysis of the probabilistic disease spread process using linear approximations and derive the convergence conditions under non-reinfection epidemic models. Finally, we conduct extensive experiments on six real-world datasets, demonstrating our method's effectiveness and robustness in estimating the nodes' final state distribution. Specifically, our proposed method consistently produces accurate estimates aligned with results from a large number of MC simulations, while maintaining a runtime comparable to a single MC simulation. Our code and datasets are available at https://github.com/GuanghuiMin/RAPID.

We study reinforcement learning from human feedback in general Markov decision processes, where agents learn from trajectory-level preference comparisons. A central challenge in this setting is to design algorithms that select informative preference queries to identify the underlying reward while ensuring theoretical guarantees. We propose a meta-algorithm based on randomized exploration, which avoids the computational challenges associated with optimistic approaches and remains tractable. We establish both regret and last-iterate guarantees under mild reinforcement learning oracle assumptions. To improve query complexity, we introduce and analyze an improved algorithm that collects batches of trajectory pairs and applies optimal experimental design to select informative comparison queries. The batch structure also enables parallelization of preference queries, which is relevant in practical deployment as feedback can be gathered concurrently. Empirical evaluation confirms that the proposed method is competitive with reward-based reinforcement learning while requiring a small number of preference queries.

Video anomaly detection (VAD) is a critical yet challenging task due to the complex and diverse nature of real-world scenarios. Previous methods typically rely on domain-specific training data and manual adjustments when applying to new scenarios and unseen anomaly types, suffering from high labor costs and limited generalization. Therefore, we aim to achieve generalist VAD, \ie, automatically handle any scene and any anomaly types without training data or human involvement. In this work, we propose PANDA, an agentic AI engineer based on MLLMs. Specifically, we achieve PANDA by comprehensively devising four key capabilities: (1) self-adaptive scene-aware strategy planning, (2) goal-driven heuristic reasoning, (3) tool-augmented self-reflection, and (4) self-improving chain-of-memory. Concretely, we develop a self-adaptive scene-aware RAG mechanism, enabling PANDA to retrieve anomaly-specific knowledge for anomaly detection strategy planning. Next, we introduce a latent anomaly-guided heuristic prompt strategy to enhance reasoning precision. Furthermore, PANDA employs a progressive reflection mechanism alongside a suite of context-aware tools to iteratively refine decision-making in complex scenarios. Finally, a chain-of-memory mechanism enables PANDA to leverage historical experiences for continual performance improvement. Extensive experiments demonstrate that PANDA achieves state-of-the-art performance in multi-scenario, open-set, and complex scenario settings without training and manual involvement, validating its generalizable and robust anomaly detection capability. Code is released at https://github.com/showlab/PANDA.

Video reasoning, the task of enabling machines to infer from dynamic visual content through multi-step logic, is crucial for advanced AI. While the Chain-of-Thought (CoT) mechanism has enhanced reasoning in text-based tasks, its application to video understanding remains underexplored. This paper presents a systematic analysis revealing that CoT often degrades performance in video reasoning, generating verbose but misleading internal monologues, and leading to hallucinated visual details and overridden correct intuitions—a phenomenon we term "visual thinking drift." We explain this drift through a Bayesian lens, positing that CoT traces often diverge from actual visual evidence, instead amplifying internal biases or language priors, causing models to storytell rather than engage in grounded reasoning. To counteract this, we introduce Visual Evidence Reward (VER), a novel reinforcement learning framework that explicitly rewards the generation of reasoning traces that are verifiably grounded in visual evidence. Comprehensive evaluation across 10 diverse video understanding benchmarks demonstrates that our Video-VER model consistently achieves top performance. Our work sheds light on the distinct challenges of video-centric reasoning and encourages the development of AI that robustly grounds its inferences in visual evidence---for large multimodal models that not only "think before answering", but also "see while thinking".

Safety verification ensures that a system avoids undesired behaviour. Liveness complements safety, ensuring that the system also achieves its desired objectives. A complete specification of functional correctness must combine both safety and liveness. Proving with mathematical certainty that a system satisfies a safety property demands presenting an appropriate inductive invariant of the system, whereas proving liveness requires showing a measure of progress witnessed by a ranking function. Neural model checking has recently introduced a data-driven approach to the formal verification of reactive systems, albeit focusing on ranking functions and thus addressing liveness properties only. In this paper, we extend and generalise neural model checking to additionally encompass inductive invariants and thus safety properties as well. Given a system and a linear temporal logic specification of safety and liveness, our approach alternates a learning and a checking component towards the construction of a provably sound neural certificate. Our new method introduces a neural certificate architecture that jointly represents inductive invariants as proofs of safety, and ranking functions as proofs of liveness. Moreover, our new architecture is amenable to training using constraint solvers, accelerating prior neural model checking work otherwise based on gradient descent. We experimentally demonstrate that our method is orders of magnitude faster than the state-of-the-art model checkers on pure liveness and combined safety and liveness verification tasks written in SystemVerilog, while enabling the verification of richer properties than was previously possible for neural model checking.

Predicting accurate future trajectories of pedestrians is essential for autonomous systems but remains a challenging task due to the need for adaptability in different environments and domains. A common approach involves collecting scenario-specific data and performing fine-tuning via backpropagation. However, the need to fine-tune for each new scenario is often impractical for deployment on edge devices. To address this challenge, we introduce TrajICL, an In-Context Learning (ICL) framework for pedestrian trajectory prediction that enables adaptation without fine-tuning on the scenario-specific data at inference time without requiring weight updates. We propose a spatio-temporal similarity-based example selection (STES) method that selects relevant examples from previously observed trajectories within the same scene by identifying similar motion patterns at corresponding locations. To further refine this selection, we introduce prediction-guided example selection (PG-ES), which selects examples based on both the past trajectory and the predicted future trajectory, rather than relying solely on the past trajectory. This approach allows the model to account for long-term dynamics when selecting examples. Finally, instead of relying on small real-world datasets with limited scenario diversity, we train our model on a large-scale synthetic dataset to enhance its prediction ability by leveraging in-context examples. Extensive experiments demonstrate that TrajICL achieves remarkable adaptation across both in-domain and cross-domain scenarios, outperforming even fine-tuned approaches across multiple public benchmarks.

Recent advances in large language models (LLMs) have intensified efforts to fuse heterogeneous open-source models into a unified system that inherits their complementary strengths. Existing logit-based fusion methods maintain inference efficiency but treat vocabulary dimensions independently, overlooking semantic dependencies encoded by cross-dimension interactions. These dependencies reflect how token types interact under a model's internal reasoning and are essential for aligning models with diverse generation behaviors. To explicitly model these dependencies, we propose \textbf{InfiGFusion}, the first structure-aware fusion framework with a novel \textit{Graph-on-Logits Distillation} (GLD) loss. Specifically, we retain the top-$k$ logits per output and aggregate their outer products across sequence positions to form a global co-activation graph, where nodes represent vocabulary channels and edges quantify their joint activations. To ensure scalability and efficiency, we design a sorting-based closed-form approximation that reduces the original $O(n^4)$ cost of Gromov-Wasserstein distance to $O(n \log n)$, with provable approximation guarantees. Experiments across multiple fusion settings show that GLD consistently improves fusion quality and stability. InfiGFusion outperforms SOTA models and fusion baselines across 11 benchmarks spanning reasoning, coding, and mathematics. It shows particular strength in complex reasoning tasks, with +35.6 improvement on Multistep Arithmetic and +37.06 on Causal Judgement over SFT, demonstrating superior multi-step and relational inference.

This study presents a new high-fidelity multi-modal dataset containing 16000+ geometric variants of automotive hoods useful for machine learning (ML) applications such as engineering component design and process optimization, and multiphysics system surrogates. The dataset is centered on a practical multiphysics problem—hood deformation from fluid entrapment and inertial loading during rotary‑dip painting. Each hood is numerically modeled with a coupled Large-Eddy Simulation (LES)-Finite Element Analysis (FEA), using 1.2M cells in total to ensure spatial and temporal accuracy. The dataset provides time-resolved physical fields, along with STL meshes and structured natural language prompts for text-to-geometry synthesis. Existing datasets are either confined to 2D cases, exhibit limited geometric variations, or lack the multi‑modal annotations and data structures—shortcomings we address with AutoHood3D. We validate our numerical methodology, establish quantitative baselines across five neural architectures, and demonstrate systematic surrogate errors in displacement and force predictions. These findings motivate the design of novel approaches and multiphysics loss functions that enforce fluid–solid coupling during model training. By providing fully reproducible workflows, AutoHood3D enables physics‑aware ML development, accelerates generative‑design iteration, and facilitates the creation of new FSI benchmarks.

This paper addresses the problem of multi-risk measure agnostic multi-armed bandits in heavy-tailed reward settings. We propose a framework that leverages novel deviation inequalities for the $1$-Wasserstein distance to construct confidence intervals for Lipschitz risk measures. The distributional LCB (DistLCB) algorithm is introduced, which achieves asymptotic optimality by deriving the first lower bounds for risk measure aware bandits with explicit sub-optimality gap dependencies. The DistLCB is further extended to multi-risk objectives, which enables Pareto-optimal solutions that consider multiple aspects of reward distributions. Additionally, we provide a regret analysis that includes both gap-dependent and gap-independent bounds for multi-risk settings. Experiments validate the effectiveness of the proposed methods in synthetic and real-world applications.

Transformers have driven remarkable breakthroughs in natural language processing and computer vision, yet their standard attention mechanism still imposes $O(N^2)$ complexity, hindering scalability to longer sequences. We introduce Circular-convolutional ATtention (CAT), a Fourier-based approach that efficiently applies circular convolutions to reduce complexity without sacrificing representational power. CAT achieves $O(N \log N)$ computations, requires fewer learnable parameters by streamlining fully connected layers, and introduces no heavier operations, resulting in consistent accuracy improvements and about a 10\% speedup in naive PyTorch implementations. Based on the engineering-isomorphic transformer framework, CAT's design not only offers practical efficiency and ease of implementation, but also provides insights to guide the development of future high-performance Transformer architectures. Finally, our ablation studies highlight the key conditions underlying CAT’s success, shedding light on broader principles for scalable attention mechanisms.

Humans excel at performing complex tasks by leveraging long-term memory across temporal and spatial experiences. In contrast, current Large Language Models (LLMs) struggle to effectively plan and act in dynamic, multi-room 3D environments. We posit that part of this limitation is due to the lack of proper 3D spatial-temporal memory modeling in LLMs. To address this, we first introduce 3DMem-Bench, a comprehensive benchmark comprising over 26,000 trajectories and 2,892 embodied tasks, question-answering and captioning, designed to evaluate an agent's ability to reason over long-term memory in 3D environments. Second, we propose 3DLLM-Mem, a novel dynamic memory management and fusion model for embodied spatial-temporal reasoning and actions in LLMs. Our model uses working memory tokens, which represents current observations, as queries to selectively attend to and fuse the most useful spatial and temporal features from episodic memory, which stores past observations and interactions. Our approach allows the agent to focus on task-relevant information while maintaining memory efficiency in complex, long-horizon environments. Experimental results demonstrate that 3DLLM-Mem achieves state-of-the-art performance across various tasks, outperforming the strongest baselines by 16.5\% in success rate on 3DMem-Bench's most challenging in-the-wild embodied tasks.