NeurIPS 2025 Papers

Safe Reinforcement Learning (RL) is crucial for achieving high performance while ensuring safety in real-world applications. However, the complex interplay of multiple uncertainty sources in real environments poses significant challenges for interpretable risk assessment and robust decision-making. To address these challenges, we propose \textbf{Fuz-RL}, a fuzzy measure-guided robust framework for safe RL. Specifically, our framework develops a novel fuzzy Bellman operator for estimating robust value functions using Choquet integrals. Theoretically, we prove that solving the Fuz-RL problem (in Constrained Markov Decision Process (CMDP) form) is equivalent to solving distributionally robust safe RL problems (in robust CMDP form), effectively avoiding min-max optimization. Empirical analyses on safe-control-gym and safety-gymnasium scenarios demonstrate that Fuz-RL effectively integrates with existing safe RL baselines in a model-free manner, significantly improving both safety and control performance under various types of uncertainties in observation, action, and dynamics.

Recent advances in multimodal Reward Models (RMs) have shown significant promise in delivering reward signals to align vision models with human preferences. However, current RMs are generally restricted to providing direct responses or engaging in shallow reasoning processes with limited depth, often leading to inaccurate reward signals. We posit that incorporating explicit long chains of thought (CoT) into the reward reasoning process can significantly strengthen their reliability and robustness. Furthermore, we believe that once RMs internalize CoT reasoning, their direct response accuracy can also be improved through implicit reasoning capabilities. To this end, this paper proposes UnifiedReward-Think, the first unified multimodal CoT-based reward model, capable of multi-dimensional, step-by-step long-chain reasoning for both visual understanding and generation reward tasks. Specifically, we adopt an exploration-driven reinforcement fine-tuning approach to elicit and incentivize the model's latent complex reasoning ability: (1) We first use a small amount of image generation preference data to distill the reasoning process of GPT-4o, which is then used for the model's cold start to learn the format and structure of CoT reasoning. (2) Subsequently, by leveraging the model's prior knowledge and generalization capabilities, we prepare large-scale unified multimodal preference data to elicit the model's reasoning process across various vision tasks. During this phase, correct reasoning outputs are retained for rejection sampling to refine the model (3) while incorrect predicted samples are finally used for Group Relative Policy Optimization (GRPO) based reinforcement fine-tuning, enabling the model to explore diverse reasoning paths and optimize for correct and robust solutions. Extensive experiments confirm that incorporating long CoT reasoning significantly enhances the accuracy of reward signals. Notably, after mastering CoT reasoning, the model exhibits implicit reasoning capabilities, allowing it to surpass existing baselines even without explicit reasoning traces.

Reinforcement learning (RL) systems have countless applications, from energy-grid management to protein design. However, such real-world scenarios are often extremely difficult, combinatorial in nature, and require complex coordination between multiple agents. This level of complexity can cause even state-of-the-art RL systems, trained until convergence, to hit a performance ceiling which they are unable to break out of with zero-shot inference. Meanwhile, many digital or simulation-based applications allow for an inference phase that utilises a specific time and compute budget to explore multiple attempts before outputting a final solution. In this work, we show that such an inference phase employed at execution time, and the choice of a corresponding inference strategy, are key to breaking the performance ceiling observed in complex multi-agent RL problems. Our main result is striking: we can obtain up to a 126% and, on average, a 45% improvement over the previous state-of-the-art across 17 tasks, using only a couple seconds of extra wall-clock time during execution. We also demonstrate promising compute scaling properties, supported by over 60k experiments, making it the largest study on inference strategies for complex RL to date. We make all of our experimental data and code available.

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/

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.

Reinforcement learning (RL) often relies on manually designed reward functions, which are difficult to specify and can lead to issues such as reward hacking and suboptimal behavior. Alternatives like inverse RL and preference-based RL attempt to infer surrogate rewards from demonstrations or preferences but suffer from ambiguity and distribution mismatch. A more direct approach, inspired by imitation learning, avoids reward modeling by leveraging expert demonstrations. However, most existing methods align actions only at individual states, failing to capture the coherence of long-horizon trajectories. In this work, we study the problem of directly aligning policies with expert-labeled trajectories to preserve long-horizon behavior without relying on reward signals. Specifically, we aim to learn a policy that maximizes the probability of generating the expert trajectories. Nevertheless, we prove that, in its general form, this trajectory alignment problem is NP-complete. To address this, we propose Trajectory Graph Learning (TGL), a framework that leverages structural assumptions commonly satisfied in practice—such as bounded realizability of expert trajectories or a tree-structured MDP. These enable a graph-based policy planning algorithm that computes optimal policies in polynomial time under known dynamics. For settings with unknown dynamics, we develop a sample-efficient algorithm based on UCB-style exploration and establish sub-linear regret. Experiments on grid-world tasks demonstrate that TGL substantially outperforms standard imitation learning methods for long-trajectory planning.

Approaches for goal-conditioned reinforcement learning (GCRL) often use learned state representations to extract goal-reaching policies. Two frameworks for representation structure have yielded particularly effective GCRL algorithms: (1) *contrastive representations*, in which methods learn "successor features" with a contrastive objective that performs inference over future outcomes, and (2) *temporal distances*, which link the (quasimetric) distance in representation space to the transit time from states to goals. We propose an approach that unifies these two frameworks, using the structure of a quasimetric representation space (triangle inequality) with the right additional constraints to learn successor representations that enable optimal goal-reaching. Unlike past work, our approach is able to exploit a **quasimetric** distance parameterization to learn **optimal** goal-reaching distances, even with **suboptimal** data and in **stochastic** environments. This gives us the best of both worlds: we retain the stability and long-horizon capabilities of Monte Carlo contrastive RL methods, while getting the free stitching capabilities of quasimetric network parameterizations. On existing offline GCRL benchmarks, our representation learning objective improves performance on stitching tasks where methods based on contrastive learning struggle, and on noisy, high-dimensional environments where methods based on quasimetric networks struggle.

Offline reinforcement learning in high-dimensional, discrete action spaces is challenging due to the exponential scaling of the joint action space with the number of sub-actions and the complexity of modeling sub-action dependencies. Existing methods either exhaustively evaluate the action space, making them computationally infeasible, or factorize Q-values, failing to represent joint sub-action effects. We propose \textbf{Bra}nch \textbf{V}alue \textbf{E}stimation (BraVE), a value-based method that uses tree-structured action traversal to evaluate a linear number of joint actions while preserving dependency structure. BraVE outperforms prior offline RL methods by up to $20\times$ in environments with over four million actions.

Value-based algorithms are a cornerstone of off-policy reinforcement learning due to their simplicity and training stability. However, their use has traditionally been restricted to discrete action spaces, as they rely on estimating Q-values for individual state-action pairs. In continuous action spaces, evaluating the Q-value over the entire action space becomes computationally infeasible. To address this, actor-critic methods are typically employed, where a critic is trained on off-policy data to estimate Q-values, and an actor is trained to maximize the critic's output. Despite their popularity, these methods often suffer from instability during training. In this work, we propose a purely value-based framework for continuous control that revisits structural maximization of Q-functions, introducing a set of key architectural and algorithmic choices to enable efficient and stable learning. We evaluate the proposed actor-free Q-learning approach on a range of standard simulation tasks, demonstrating performance and sample-efficiency on par with state-of-the-art baselines, without the cost of learning a separate actor. Particularly, in environments with constrained action spaces, where the value functions are typically non-smooth, our method with structural maximization outperforms traditional actor-critic methods with gradient-based maximization.

Symmetry is pervasive in robotics and has been widely exploited to improve sample efficiency in deep reinforcement learning (DRL). However, existing approaches primarily focus on spatial symmetries—such as reflection, rotation, and translation—while largely neglecting temporal symmetries. To address this gap, we explore time reversal symmetry, a form of temporal symmetry commonly found in robotics tasks such as door opening and closing. We propose Time Reversal symmetry enhanced Deep Reinforcement Learning (TR-DRL), a framework that combines trajectory reversal augmentation and time reversal guided reward shaping to efficiently solve temporally symmetric tasks. Our method generates reversed transitions from fully reversible transitions, identified by a proposed dynamics-consistent filter, to augment the training data. For partially reversible transitions, we apply reward shaping to guide learning, according to successful trajectories from the reversed task. Extensive experiments on the Robosuite and MetaWorld benchmarks demonstrate that TR-DRL is effective in both single-task and multi-task settings, achieving higher sample efficiency and stronger final performance compared to baseline methods.

Establishing stability certificates for closed-loop systems under reinforcement learning (RL) policies is essential to move beyond empirical performance and offer guarantees of system behavior. Classical Lyapunov methods require a strict stepwise decrease in the Lyapunov function but such certificates are difficult to construct for learned policies. The RL value function is a natural candidate but it is not well understood how it can be adapted for this purpose. To gain intuition, we first study the linear quadratic regulator (LQR) problem and make two key observations. First, a Lyapunov function can be obtained from the value function of an LQR policy by augmenting it with a residual term related to the system dynamics and stage cost. Second, the classical Lyapunov decrease requirement can be relaxed to a generalized Lyapunov condition requiring only decrease on average over multiple time steps. Using this intuition, we consider the nonlinear setting and formulate an approach to learn generalized Lyapunov functions by augmenting RL value functions with neural network residual terms. Our approach successfully certifies the stability of RL policies trained on Gymnasium and DeepMind Control benchmarks. We also extend our method to jointly train neural controllers and stability certificates using a multi-step Lyapunov loss, resulting in larger certified inner approximations of the region of attraction compared to the classical Lyapunov approach. Overall, our formulation enables stability certification for a broad class of systems with learned policies by making certificates easier to construct, thereby bridging classical control theory and modern learning-based methods.

Given the ever-changing nature of the world and its inhabitants, agents must possess the ability to adapt and evolve over time. Recent research in Given the ever-changing nature of the world and its inhabitants, agents must possess the ability to adapt and evolve over time. Recent research in non-stationary MDPs has focused on addressing this challenge, providing algorithms inspired by task inference techniques. However, these methods ignore the detrimental effects of interference, which particularly harm performance in contradictory tasks, leading to low efficiency in some environments. To address this issue, we propose a Bayesian Fast-Slow Framework (BFSF) that tackles both cross-task generalization and resistance to cross-task interference. Our framework consists of two components: a 'fast' policy, learned from recent data, and a 'slow' policy, learned through meta-reinforcement learning (meta-RL) using data from all previous tasks. A Bayesian estimation mechanism determines the current choice of 'fast' or 'slow' policy, balancing exploration and exploitation. Additionally, in the 'fast' policy, we introduce a dual-reset mechanism and a data relabeling technique to further accelerate convergence when encountering new tasks. Experiments demonstrate that our algorithm effectively mitigates interference and outperforms baseline approaches.

Designing reinforcement learning agents to satisfy complex temporal objectives expressed in Linear Temporal Logic (LTL), presents significant challenges, particularly in ensuring sample efficiency and task alignment over infinite horizons. Recent works have shown that by leveraging the corresponding Limit Deterministic Büchi Automaton (LDBA) representation, LTL formulas can be translated into variable discounting schemes over LDBA-accepting states to maximize a lower bound on the probability of formula satisfaction. However, the resulting reward signals are inherently sparse, making exploration of LDBA-accepting states increasingly difficult as task horizons lengthen to infinity. In this work, we address these challenges by leveraging finite-length demonstrations to overcome the exploration bottleneck for LTL objectives over infinite horizons. We segment demonstrations and agent exploratory trajectories at LDBA-accepting states and iteratively guide the agent within each segment to learn to reach these accepting states. By incentivizing the agent to visit LDBA-accepting states from arbitrary states, our approach increases the probability of LTL formula satisfaction without the need for extensive or lengthy demonstrations. We demonstrate the applicability of our method across a variety of high-dimensional continuous control domains. It achieves faster convergence and consistently outperforms baseline approaches.

Scaling deep reinforcement learning in pixel-based environments presents a significant challenge, often resulting in diminished performance. While recent works have proposed algorithmic and architectural approaches to address this, the underlying cause of the performance drop remains unclear. In this paper, we identify the connection between the output of the encoder (a stack of convolutional layers) and the ensuing dense layers as the main underlying factor limiting scaling capabilities; we denote this connection as the **bottleneck**, and we demonstrate that previous approaches implicitly target this bottleneck. As a result of our analyses, we present global average pooling as a simple yet effective way of targeting the bottleneck, thereby avoiding the complexity of earlier approaches.

Online advertising platforms use automated auctions to connect advertisers with potential customers, requiring effective bidding strategies to maximize profits. Accurate ad impact estimation requires considering three key factors: delayed and long-term effects, cumulative ad impacts such as reinforcement or fatigue, and customer heterogeneity. However, these effects are often not jointly addressed in previous studies. To capture these factors, we model ad bidding as a Contextual Markov Decision Process (CMDP) with delayed Poisson rewards. For efficient estimation, we propose a two-stage maximum likelihood estimator combined with data-splitting strategies, ensuring controlled estimation error based on the first-stage estimator's (in)accuracy. Building on this, we design a reinforcement learning algorithm to derive efficient personalized bidding strategies. This approach achieves a near-optimal regret bound of $\tilde{\mathcal{O}}(dH^2\sqrt{T})$, where $d$ is the contextual dimension, $H$ is the number of rounds, and $T$ is the number of customers. Our theoretical findings are validated by simulation experiments.

Continual offline reinforcement learning (CORL) has shown impressive ability in diffusion-based continual learning systems by modeling the joint distributions of trajectories. However, most research only focuses on limited continual task settings where the tasks have the same observation and action space, which deviates from the realistic demands of training agents in various environments. In view of this, we propose Vector-Quantized Continual Diffuser, named VQ-CD, to break the barrier of different spaces between various tasks. Specifically, our method contains two complementary sections, where the quantization spaces alignment provides a unified basis for the selective weights activation. In the quantized spaces alignment, we leverage vector quantization to align the different state and action spaces of various tasks, facilitating continual training in the same space. Then, we propose to leverage a unified diffusion model attached by the inverse dynamic model to master all tasks by selectively activating different weights according to the task-related sparse masks. Finally, we conduct extensive experiments on 15 continual learning (CL) tasks, including conventional CL task settings (identical state and action spaces) and general CL task settings (various state and action spaces). Compared with 17 baselines, our method reaches the SOTA performance.

Using a long-standing conjecture from combinatorial group theory, we explore, from multiple perspectives, the challenges of finding rare instances carrying disproportionately high rewards. Based on lessons learned in the context defined by the Andrews--Curtis conjecture, we analyze how reinforcement learning agents handle problems of varying hardness. We also address many mathematical questions as a part of our study. Notably, we demonstrate the length reducibility of all but two presentations in the Akbulut--Kirby series (1981), and resolve various potential counterexamples in the Miller--Schupp series (1991), including three infinite subfamilies.

Constructing multistage optimal decisions for alternating recurrent event data is critically important in medical and healthcare research. Current reinforcement learning (RL) algorithms have only been applied to time-to-event data, with the objective of maximizing expected survival time. However, alternating recurrent event data has a different structure, which motivates us to model the probability and frequency of event occurrences rather than a single terminal outcome. In this paper, we introduce an RL framework specifically designed for alternating recurrent event data. Our goal is to maximize the probability that the duration between consecutive events exceeds a clinically meaningful threshold. To achieve this, we identify a lower bound of this probability, which transforms the problem into maximizing a cumulative sum of log probabilities, thus enabling direct application of standard RL algorithms. We establish the theoretical properties of the resulting optimal policy and demonstrate through numerical experiments that our proposed algorithm yields a larger probability of that the time between events exceeds a critical threshold compared with existing state-of-the-art algorithms.

Mean field games (MFGs) have emerged as a powerful framework for modeling interactions in large-scale multi-agent systems. Despite recent advancements in reinforcement learning (RL) for MFGs, existing methods are typically limited to finite spaces or stationary models, hindering their applicability to real-world problems. This paper introduces a novel deep reinforcement learning (DRL) algorithm specifically designed for non-stationary continuous MFGs. The proposed approach builds upon a Fictitious Play (FP) methodology, leveraging DRL for best-response computation and supervised learning for average policy representation. Furthermore, it learns a representation of the time-dependent population distribution using a Conditional Normalizing Flow. To validate the effectiveness of our method, we evaluate it on three different examples of increasing complexity. By addressing critical limitations in scalability and density approximation, this work represents a significant advancement in applying DRL techniques to complex MFG problems, bringing the field closer to real-world multi-agent systems.