NeurIPS 2025 论文列表

Monte Carlo Tree Search (MCTS), which leverages Upper Confidence Bound for Trees (UCTs) to balance exploration and exploitation through randomized sampling, is instrumental to solving complex planning problems. However, for multi-agent planning, MCTS is confronted with a large combinatorial action space that often grows exponentially with the number of agents. As a result, the branching factor of MCTS during tree expansion also increases exponentially, making it very difficult to efficiently explore and exploit during tree search. To this end, we propose MALinZero, a new approach to leverage low-dimensional representational structures on joint-action returns and enable efficient MCTS in complex multi-agent planning. Our solution can be viewed as projecting the joint-action returns into the low-dimensional space representable using a contextual linear bandit problem formulation. We solve the contextual linear bandit problem with convex and $\mu$-smooth loss functions -- in order to place more importance on better joint actions and mitigate potential representational limitations -- and derive a linear Upper Confidence Bound applied to trees (LinUCT) to enable novel multi-agent exploration and exploitation in the low-dimensional space. We analyze the regret of MALinZero for low-dimensional reward functions and propose an $(1-\tfrac1e)$-approximation algorithm for the joint action selection by maximizing a sub-modular objective. MALinZero demonstrates state-of-the-art performance on multi-agent benchmarks such as matrix games, SMAC, and SMACv2, outperforming both model-based and model-free multi-agent reinforcement learning baselines with faster learning speed and better performance.

A robot's instantaneous sensory observations do not always reveal task-relevant state information. Under such partial observability, optimal behavior typically involves explicitly acting to gain the missing information. Today's standard robot learning techniques struggle to produce such active perception behaviors. We propose a simple real-world robot learning recipe to efficiently train active perception policies. Our approach, asymmetric advantage weighted regression (AAWR), exploits access to "privileged" extra sensors at training time. The privileged sensors enable training high-quality privileged value functions that aid in estimating the advantage of the target policy. Bootstrapping from a small number of potentially suboptimal demonstrations and an easy-to-obtain coarse policy initialization, AAWR quickly acquires active perception behaviors and boosts task performance. In evaluations on 8 manipulation tasks on 3 robots spanning varying degrees of partial observability, AAWR synthesizes reliable active perception behaviors that outperform all prior approaches. When initialized with a "generalist" robot policy that struggles with active perception tasks, AAWR efficiently generates information-gathering behaviors that allow it to operate under severe partial observability for manipulation tasks. Website: https://penn-pal-lab.github.io/aawr/

Visual geo-localization for drones faces critical degradation under weather perturbations, \eg, rain and fog, where existing methods struggle with two inherent limitations: 1) Heavy reliance on limited weather categories that constrain generalization, and 2) Suboptimal disentanglement of entangled scene-weather features through pseudo weather categories. We present WeatherPrompt, a multi-modality learning paradigm that establishes weather-invariant representations through fusing the image embedding with the text context. Our framework introduces two key contributions: First, a Training-free Weather Reasoning mechanism that employs off-the-shelf large multi-modality models to synthesize multi-weather textual descriptions through human-like reasoning. It improves the scalability to unseen or complex weather, and could reflect different weather strength. Second, to better disentangle the scene and weather features, we propose a multi-modality framework with the dynamic gating mechanism driven by the text embedding to adaptively reweight and fuse visual features across modalities. The framework is further optimized by the cross-modal objectives, including image-text contrastive learning and image-text matching, which maps the same scene with different weather conditions closer in the representation space. Extensive experiments validate that, under diverse weather conditions, our method achieves competitive recall rates compared to state-of-the-art drone geo-localization methods. Notably, it improves Recall@1 by 13.37\% under night conditions and by 18.69\% under fog and snow conditions. Our code is available at https://github.com/Jahawn-Wen/WeatherPrompt.

Video Temporal Grounding (TG) aims to temporally locate video segments matching a natural language description (a query) in a long video. While Vision-Language Models (VLMs) are effective at holistic semantic matching, they often struggle with fine-grained temporal localisation. Recently, Group Relative Policy Optimisation (GRPO) reformulates the inference process as a reinforcement learning task, enabling fine-grained grounding and achieving strong in-domain performance. However, GRPO relies on labelled data, making it unsuitable in unlabelled domains. Moreover, because videos are large and expensive to store and process, performing full-scale adaptation introduces prohibitive latency and computational overhead, making it impractical for real-time deployment. To overcome both problems, we introduce a Data-Efficient Unlabelled Cross-domain Temporal Grounding method, from which a model is first trained on a labelled source domain, then adapted to a target domain using only a small number of {\em unlabelled videos from the target domain}. This approach eliminates the need for target annotation and keeps both computational and storage overhead low enough to run in real time. Specifically, we introduce \textbf{U}ncertainty-quantified \textbf{R}ollout \textbf{P}olicy \textbf{A}daptation (\textbf{URPA}) for cross-domain knowledge transfer in learning video temporal grounding without target labels. URPA generates multiple candidate predictions using GRPO rollouts, averages them to form a pseudo label, and estimates confidence from the variance across these rollouts. This confidence then weights the training rewards, guiding the model to focus on reliable supervision. Experiments on three datasets across six cross-domain settings show that URPA generalises well using only a few unlabelled target videos. Codes are given in supplemental materials.

We take two key steps in automating the open-ended discovery of new mathematical theories, a grand challenge in artificial intelligence. First, we introduce Fermat, a reinforcement learning (RL) environment that models concept discovery and theorem-proving using a set of symbolic actions, opening up a range of RL problems relevant to theory discovery. Second, we explore a specific problem through Fermat: automatically scoring the interestingness of mathematical objects. We investigate evolutionary algorithms for synthesizing nontrivial interestingness measures. In particular, we introduce an LLM-based evolutionary algorithm that features function abstraction, leading to notable improvements in discovering elementary number theory and finite fields over hard-coded baselines. We open-source the \fermat environment at github.com/trishullab/Fermat.

Large Language Models (LLMs) face significant limitations when applied to large-scale graphs, struggling with context constraints and inflexible reasoning. We introduce GraphChain, a novel framework enabling LLMs to analyze large graphs by orchestrating dynamic sequences of specialized tools, mimicking human exploratory processes. GraphChain incorporates two core technical contributions: (1) Progressive Graph Distillation, a reinforcement learning approach that learns to generate tool sequences balancing task relevance and intermediate state compression, thereby overcoming LLM context limitations. (2) Structure-aware Test-Time Adaptation (STTA), a mechanism using a lightweight, self-supervised adapter conditioned on graph spectral properties to efficiently adapt a frozen LLM policy to diverse graph structures via soft prompts without retraining. Experiments show GraphChain significantly outperforms prior methods, enabling scalable and adaptive LLM-driven graph analysis.

Learning from multi-agent expert demonstrations, known as Multi-Agent Imitation Learning (MAIL), provides a promising approach to sequential decision-making. However, existing MAIL methods including Behavior Cloning (BC) and Adversarial Imitation Learning (AIL) face significant challenges: BC suffers from the compounding error issue, while the very nature of adversarial optimization makes AIL prone to instability. In this work, we propose \textbf{M}ulti-\textbf{A}gent imitation by learning and sampling from \textbf{F}actor\textbf{I}zed \textbf{S}oft Q-function (MAFIS), a novel method that addresses these limitations for both online and offline MAIL settings. Built upon the single-agent IQ-Learn framework, MAFIS introduces the value decomposition network to factorize the imitation objective at agent level, thus enabling scalable training for multi-agent systems. Moreover, we observe that the soft Q-function implicitly defines the optimal policy as an energy-based model, from which we can sample actions via stochastic gradient Langevin dynamics. This allows us to estimate the gradient of the factorized optimization objective for continuous control tasks, avoiding the adversarial optimization between the soft Q-function and the policy required by prior work. By doing so, we obtain a tractable and \emph{non-adversarial} objective for both discrete and continuous multi-agent control. Experiments on common benchmarks including the discrete control tasks StarCraft Multi-Agent Challenge v2 (SMACv2), Gold Miner, and Multi Particle Environments (MPE), as well as the continuous control task Multi-Agent MuJoCo (MaMuJoCo), demonstrate that MAFIS achieves superior performance compared with baselines. Our code is available at https://github.com/LAMDA-RL/MAFIS.

Recent studies have demonstrated the efficacy of using Reinforcement Learning (RL) in building reasoning models that articulate chains of thoughts prior to producing final answers. However, despite ongoing advances that aim at enabling reasoning for vision-language tasks, existing open-source visual reasoning models typically generate reasoning content with pure natural language, lacking explicit integration of visual information. This limits their ability to produce clearly articulated and visually grounded reasoning chains. To this end, we propose Grounded Reasoning with Images and Texts (GRIT), a novel method for training MLLMs to think with images. GRIT introduces a grounded reasoning paradigm, in which models generate reasoning chains that interleave natural language and explicit bounding box coordinates. These coordinates point to regions of the input image that the model consults during its reasoning process. Additionally, GRIT is equipped with a reinforcement learning approach, GRPO-GR, built upon the GRPO algorithm. GRPO-GR employs robust rewards focused on the final answer accuracy and format of the grounded reasoning output, which eliminates the need for data with reasoning chain annotations or explicit bounding box labels. As a result, GRIT achieves exceptional data efficiency, requiring as few as 20 image-question-answer triplets from existing datasets. Comprehensive evaluations demonstrate that GRIT effectively trains MLLMs to produce coherent and visually grounded reasoning chains, showing a successful unification of reasoning and grounding abilities. All code, data, and checkpoints will be released.

The rapid progress of large language models (LLMs) has catalyzed the emergence of multimodal large language models (MLLMs) that unify visual understanding and image generation within a single framework. However, most existing MLLMs rely on autoregressive (AR) architectures, which impose inherent limitations on future development, such as the raster-scan order in image generation and restricted reasoning abilities in causal context modeling. In this work, we challenge the dominance of AR-based approaches by introducing FUDOKI, a unified multimodal model purely based on discrete flow matching, as an alternative to conventional AR paradigms. By leveraging metric-induced probability paths with kinetic optimal velocities, our framework goes beyond the previous masking-based corruption process, enabling iterative refinement with self-correction capability and richer bidirectional context integration during generation. To mitigate the high cost of training from scratch, we initialize FUDOKI from pre-trained AR-based MLLMs and adaptively transition to the discrete flow matching paradigm. Experimental results show that FUDOKI achieves performance comparable to state-of-the-art AR-based MLLMs across both visual understanding and image generation tasks, highlighting its potential as a foundation for next-generation unified multimodal models. Furthermore, we show that applying test-time scaling techniques to FUDOKI yields significant performance gains, further underscoring its promise for future enhancement through reinforcement learning.

Spiking Neural Networks (SNNs) offer low-latency and energy-efficient decision making on neuromorphic hardware, making them attractive for Reinforcement Learning (RL) in resource-constrained edge devices. However, most RL algorithms for continuous control are designed for Artificial Neural Networks (ANNs), particularly the target network soft update mechanism, which conflicts with the discrete and non-differentiable dynamics of spiking neurons. We show that this mismatch destabilizes SNN training and degrades performance. To bridge the gap between discrete SNNs and continuous-control algorithms, we propose a novel proxy target framework. The proxy network introduces continuous and differentiable dynamics that enable smooth target updates, stabilizing the learning process. Since the proxy operates only during training, the deployed SNN remains fully energy-efficient with no additional inference overhead. Extensive experiments on continuous control benchmarks demonstrate that our framework consistently improves stability and achieves up to $32$% higher performance across various spiking neuron models. Notably, to the best of our knowledge, this is the first approach that enables SNNs with simple Leaky Integrate and Fire (LIF) neurons to surpass their ANN counterparts in continuous control. This work highlights the importance of SNN-tailored RL algorithms and paves the way for neuromorphic agents that combine high performance with low power consumption. Code is available at https://github.com/xuzijie32/Proxy-Target.

Language models must be adapted to understand and follow user instructions. Reinforcement learning is widely used to facilitate this —typically using fixed criteria such as "helpfulness" and "harmfulness". In our work, we instead propose using flexible, instruction-specific criteria as a means of broadening the impact that reinforcement learning can have in eliciting instruction following. We propose "Reinforcement Learning from Checklist Feedback" (RLCF). From instructions, we extract checklists and evaluate how well responses satisfy each item—using both AI judges and specialized verifier programs—then combine these scores to compute rewards for RL. We compare RLCF with other alignment methods on top of a strong instruction following model (Qwen2.5-7B-Instruct) on five widely-studied benchmarks — RLCF is the only method to help on every benchmark, including a 4-point boost in hard satisfaction rate on FollowBench, a 6-point increase on InFoBench, and a 3-point rise in win rate on Arena-Hard. We show that RLCF can also be used off-policy to improve Llama 3.1 8B Instruct and OLMo 2 7B Instruct. These results establish rubrics as a key tool for improving language models' support of queries that express a multitude of needs. We release our our dataset of rubrics (WildChecklists), models, and code to the public.

Reliable uncertainty quantification is crucial for reinforcement learning (RL) in high-stakes settings. We propose a unified conformal prediction framework for infinite-horizon policy evaluation that constructs distribution-free prediction intervals for returns in both on-policy and off-policy settings. Our method integrates distributional RL with conformal calibration, addressing challenges such as unobserved returns, temporal dependencies, and distributional shifts. We propose a modular pseudo-return construction based on truncated rollouts and a time-aware calibration strategy using experience replay and weighted subsampling. These innovations mitigate model bias and restore approximate exchangeability, enabling uncertainty quantification even under policy shifts. Our theoretical analysis provides coverage guarantees that account for model misspecification and importance weight estimation. Empirical results, including experiments in synthetic and benchmark environments like Mountain Car, show that our method significantly improves coverage and reliability over standard distributional RL baselines.

Large language models can generate rare but catastrophic outputs, such as harmful conversations or insecure code. Existing Reinforcement Learning from Human Feedback (RLHF) typically maximizes average reward, leaving high-risk tail events insufficiently controlled. We introduce Quantile‑Guided Alignment (QA), a framework that allows users to specify desired improvements at any quantile—individually or across multiple reward dimensions—thus shifting the distribution of outputs with finer control toward safer, more desirable outcomes. The method extends standard RLHF via an augmented reward formulation that enforces quantile constraints. Experiments on conversation and code‐generation tasks show that quantile alignment significantly enhances quality at targeted tails while maintaining overall performance. The results position QA as a principled route to risk‑calibrated language models with tail‑focused alignment.

Recent advancements in vision-language models (VLMs) have improved performance by increasing the number of visual tokens, which are often significantly longer than text tokens. However, we observe that most real-world scenarios do not require such an extensive number of visual tokens. While the performance drops significantly in a small subset of OCR-related tasks, models still perform accurately in most other general VQA tasks with only 1/4 resolution. Therefore, we propose to dynamically process distinct samples with different resolutions, and present a new paradigm for visual token compression, namely, VisionThink. It starts with a downsampled image and smartly decides whether it is sufficient for problem solving. Otherwise, the model could output a special token to request the higher-resolution image. Compared to existing Efficient VLM methods that compress tokens using fixed pruning ratios or thresholds, VisionThink autonomously decides whether to compress tokens case by case. As a result, it demonstrates strong fine-grained visual understanding capability on OCR-related tasks, and meanwhile saves substantial visual tokens on simpler tasks. We adopt reinforcement learning and propose the LLM-as-Judge strategy to successfully apply RL to general VQA tasks. Moreoever, we carefully design a reward function and penalty mechanism to achieve a stable and reasonable image resize call ratio. Extensive experiments demonstrate the superiority, efficiency, and effectiveness of our method. All our code and data are open-sourced.

Emergency traffic control presents an increasingly critical challenge, requiring seamless coordination among emergency vehicles, regular vehicles, and traffic lights to ensure efficient passage for all vehicles. Existing models primarily only focus on traffic light control, leaving emergency and regular vehicles prone to delay due to the lack of navigation strategies. To address this issue, we propose the ***R*ole-aware *M*ulti-agent *T*raffic *C*ontrol (RMTC)** framework, which dynamically assigns appropriate roles to traffic components for better cooperation by considering their relations with emergency vehicles and adaptively adjusting their policies. Specifically, RMTC introduces a *Heterogeneous Temporal Traffic Graph (HTTG)* to model the spatial and temporal relationships among all traffic components (traffic lights, regular and emergency vehicles) at each time step. Furthermore, we develop a *Dynamic Role Learning* model to infer the evolving roles of traffic lights and regular vehicles based on HTTG. Finally, we present a *Role-aware Multi-agent Reinforcement Learning* approach that learns traffic policies conditioned on the dynamically roles. Extensive experiments across four public traffic scenarios show that RMTC outperforms existing traffic light control methods by significantly reducing emergency vehicle travel time, while effectively preserving traffic efficiency for regular vehicles. The code is released at [https://anonymous.4open.science/r/RMTC-5E28](https://anonymous.4open.science/r/RMTC-5E28).

In modern sequential decision-making systems, the construction of an optimal candidate action space is critical to efficient inference. However, existing approaches either rely on manually defined action spaces that lack scalability or utilize unstructured spaces that render exhaustive search computationally prohibitive. In this paper, we propose a novel framework named \textsc{DynaAct} for automatically constructing a compact action space to enhance sequential reasoning in complex problem-solving scenarios. Our method first estimates a proxy for the complete action space by extracting general sketches observed in a corpus covering diverse complex reasoning problems using large language models. We then formulate a submodular function that jointly evaluates candidate actions based on their utility to the current state and their diversity, and employ a greedy algorithm to select an optimal candidate set. Extensive experiments on six diverse standard benchmarks demonstrate that our approach significantly improves overall performance, while maintaining efficient inference without introducing substantial latency. The implementation is available at \url{https://github.com/zhaoxlpku/DynaAct}.

Bisimulation metrics are powerful tools for measuring similarities between stochastic processes, and specifically Markov chains. Recent advances have uncovered that bisimulation metrics are, in fact, optimal-transport distances, which has enabled the development of fast algorithms for computing such metrics with provable accuracy and runtime guarantees. However, these recent methods, as well as all previously known methods, assume full knowledge of the transition dynamics. This is often an impractical assumption in most real-world scenarios, where typically only sample trajectories are available. In this work, we propose a stochastic optimization method that addresses this limitation and estimates bisimulation metrics based on sample access, without requiring explicit transition models. Our approach is derived from a new linear programming (LP) formulation of bisimulation metrics, which we solve using a stochastic primal-dual optimization method. We provide theoretical guarantees on the sample complexity of the algorithm and validate its effectiveness through a series of empirical evaluations.

The fine-tuning of pre-trained large language models (LLMs) using reinforcement learning (RL) is generally formulated as direct policy optimization. This approach was naturally favored as it efficiently improves a pretrained LLM with simple gradient updates. Another RL paradigm, Q-learning methods, has received far less attention in the LLM community while demonstrating major success in various non-LLM RL tasks. In particular, Q-learning effectiveness stems from its sample efficiency and ability to learn offline, which is particularly valuable given the high computational cost of sampling with LLM. However, naively applying a Q-learning–style update to the model’s logits is ineffective due to the specificity of LLMs. Our contribution is to derive theoretically grounded loss functions from Bellman equations to adapt Q-learning methods to LLMs. To do so, we interpret LLM logits as Q-values and carefully adapt insights from the RL literature to account for LLM-specific characteristics. It thereby ensures that the logits become reliable Q-value estimates. We then use this loss to build a practical algorithm, ShiQ for Shifted-Q, that supports off-policy, token-wise learning while remaining simple to implement. Finally, ShiQ is evaluated on both synthetic data and real-world benchmarks, e.g., UltraFeedback, BFCL-V3, demonstrating its effectiveness in both single-turn and multi-turn LLM settings.

Accurate and safe medication recommendations are critical for effective clinical decision-making, especially in multimorbidity cases. However, existing systems rely on point-wise prediction paradigms that overlook synergistic drug effects and potential adverse drug-drug interactions (DDIs). We propose FLAME, a fine-grained list-wise alignment framework for large language models (LLMs), enabling drug-by-drug generation of drug lists. FLAME formulates recommendation as a sequential decision process, where each step adds or removes a single drug. To provide fine-grained learning signals, we devise step-wise Group Relative Policy Optimization (GRPO) with potential-based reward shaping, which explicitly models DDIs and optimizes the contribution of each drug to the overall prescription. Furthermore, FLAME enhances patient modeling by integrating structured clinical knowledge and collaborative information into the representation space of LLMs. Experiments on benchmark datasets demonstrate that FLAME achieves state-of-the-art performance, delivering superior accuracy, controllable safety–accuracy trade-offs, and strong generalization across diverse clinical scenarios. Our code is available at https://github.com/cxfann/Flame.