NeurIPS 2025 论文列表

Reinforcement learning (RL) has emerged as a powerful post-training technique to incentivize the reasoning ability of large language models (LLMs). However, LLMs can respond very inconsistently to RL finetuning: some show substantial performance gains, while others plateau or even degrade. To understand this divergence, we analyze the per-step influence of the RL objective and identify two key conditions for effective post-training: (1) RL-informative rollout accuracy, and (2) strong data co-influence, which quantifies how much the training data affects performance on other samples. Guided by these insights, we propose behavior injection, a task-agnostic data augmentation scheme applied prior to RL. Behavior injection enriches the supervised finetuning (SFT) data by seeding exploratory and exploitative behaviors, effectively making the model more RL-ready. We evaluate our method across two reasoning benchmarks with multiple base models. The results demonstrate that our theoretically motivated augmentation can significantly increase the performance gain from RL over the pre-RL model.

We study online preference-based reinforcement learning (PbRL) with the goal of improving sample efficiency. While a growing body of theoretical work has emerged—motivated by PbRL’s recent empirical success, particularly in aligning large language models (LLMs)—most existing studies focus only on pairwise comparisons. A few recent works (Zhu et al., 2023, Mukherjee et al., 2024, Thekumparampil et al., 2024) have explored using multiple comparisons and ranking feedback, but their performance guarantees fail to improve—and can even deteriorate—as the feedback length increases, despite the richer information available. To address this gap, we adopt the Plackett–Luce (PL) model for ranking feedback over action subsets and propose **M-AUPO**, an algorithm that selects multiple actions by maximizing the average uncertainty within the offered subset. We prove that **M-AUPO** achieves a suboptimality gap of $\tilde{\mathcal{O}}\left( \frac{d}{T} \sqrt{ \sum_{t=1}^T \frac{1}{|S_t|}} \right)$, where $T$ is the total number of rounds, $d$ is the feature dimension, and $|S_t|$ is the size of the subset at round $t$. This result shows that larger subsets directly lead to improved performance and, notably, the bound avoids the exponential dependence on the unknown parameter’s norm, which was a fundamental limitation in most previous works. Moreover, we establish a near-matching lower bound of $\Omega \left( \frac{d}{K \sqrt{T}} \right)$, where $K$ is the maximum subset size. To the best of our knowledge, this is the first theoretical result in PbRL with ranking feedback that explicitly shows improved sample efficiency as a function of the subset size.

Low-rank structure is a common implicit assumption in many modern reinforcement learning (RL) algorithms. For instance, reward-free and goal-conditioned RL methods often presume that the successor measure admits a low-rank representation. In this work, we challenge this assumption by first remarking that the successor measure itself is not approximately low-rank. Instead, we demonstrate that a low-rank structure naturally emerges in the shifted successor measure, which captures the system dynamics after bypassing a few initial transitions. We provide finite-sample performance guarantees for the entry-wise estimation of a low-rank approximation of the shifted successor measure from sampled entries. Our analysis reveals that both the approximation and estimation errors are primarily governed by a newly introduced quantitity: the spectral recoverability of the corresponding matrix. To bound this parameter, we derive a new class of functional inequalities for Markov chains that we call Type II Poincaré inequalities and from which we can quantify the amount of shift needed for effective low-rank approximation and estimation. This analysis shows in particular that the required shift depends on decay of the high-order singular values of the shifted successor measure and is hence typically small in practice. Additionally, we establish a connection between the necessary shift and the local mixing properties of the underlying dynamical system, which provides a natural way of selecting the shift. Finally, we validate our theoretical findings with experiments, and demonstrate that shifting the successor measure indeed leads to improved performance in goal-conditioned RL.

Reinforcement learning has achieved remarkable success in complex decision-making environments, yet its lack of transparency limits its deployment in practice, especially in safety-critical settings. Shapley values from cooperative game theory provide a principled framework for explaining reinforcement learning; however, the computational cost of Shapley explanations is an obstacle for their use. We introduce FastSVERL, a scalable method for explaining reinforcement learning by approximating Shapley values. FastSVERL is designed to handle the unique challenges of reinforcement learning, including temporal dependencies across multi-step trajectories, learning from off-policy data, and adapting to evolving agent behaviours in real time. FastSVERL introduces a practical, scalable approach for principled and rigourous interpretability in reinforcement learning.

Trust in a decision-making system requires both safety guarantees and the ability to interpret and understand its behavior. This is particularly important for learned systems, whose decision-making processes are often highly opaque. Shielding is a prominent model-based technique for enforcing safety in reinforcement learning. However, because shields are automatically synthesized using rigorous formal methods, their decisions are often similarly difficult for humans to interpret. Recently, decision trees became customary to represent controllers and policies. However, since shields are inherently non-deterministic, their decision tree representations become too large to be explainable in practice. To address this challenge, we propose a novel approach for explainable safe RL that enhances trust by providing human-interpretable explanations of the shield's decisions. Our method represents the shielding policy as a hierarchy of decision trees, offering top-down, case-based explanations. At design time, we use a world model to analyze the safety risks of executing actions in given states. Based on this risk analysis, we construct both the shield and a high-level decision tree that classifies states into risk categories (safe, critical, dangerous, unsafe), providing an initial explanation of why a given situation may be safety-critical. At runtime, we generate localized decision trees that explain which actions are allowed and why others are deemed unsafe. Altogether, our method facilitates the explainability of the safety aspect in the safe-by-shielding reinforcement learning. Our framework requires no additional information beyond what is already used for shielding, incurs minimal overhead, and can be readily integrated into existing shielded RL pipelines. In our experiments, we compute explanations using decision trees that are several orders of magnitude smaller than the original shield.

World models predict state transitions in response to actions and are increasingly developed across diverse modalities. However, standard training objectives such as maximum likelihood estimation (MLE) often misalign with task-specific goals of world models, i.e., transition prediction metrics like accuracy or perceptual quality. In this paper, we present RLVR-World, a unified framework that leverages reinforcement learning with verifiable rewards (RLVR) to directly optimize world models for such metrics. Despite formulating world modeling as autoregressive prediction of tokenized sequences, RLVR-World evaluates metrics of decoded predictions as verifiable rewards. We demonstrate substantial performance gains on both language- and video-based world models across domains, including text games, web navigation, and robot manipulation. Our work indicates that, beyond recent advances in reasoning language models, RLVR offers a promising post-training paradigm for enhancing the utility of generative models more broadly. Code, datasets, models, and video samples are available at the project website: https://thuml.github.io/RLVR-World.

Reinforcement Learning from Human Feedback (RLHF) has emerged as a key technique for post‑training large language models. Despite its empirical success, the theoretical understanding of RLHF is still limited, as learning the KL-regularized target with only preference feedback poses additional challenges compared with canonical RL. Existing works mostly study the reward-based Bradley-Terry (BT) preference model, and extend classical designs utilizing optimism or pessimism. This work, instead, considers the general preference model (whose practical relevance has been observed recently) and obtains performance guarantees with major, order-wise improvements over existing ones. Surprisingly, these results are derived from algorithms that directly use empirical estimates (i.e., greedy sampling), as opposed to constructing optimistic or pessimistic estimates in previous works. This insight has a deep root in the unique structural property of the optimal policy class under the KL-regularized target, and we further specialize it to the BT model, highlighting the surprising sufficiency of greedy sampling in RLHF.

Cutting planes (cuts) are essential for solving mixed-integer linear programming (MILP) problems, as they tighten the feasible solution space and accelerate the solving process. Modern MILP solvers offer diverse cutting plane separators to generate cuts, enabling users to leverage their potential complementary strengths to tackle problems with different structures. Recent machine learning approaches learn to configure separators based on problem-specific features, selecting effective separators and deactivating ineffective ones to save unnecessary computing time. However, they ignore the dynamics of separator efficacy at different stages of cut generation and struggle to adapt the configurations for the evolving problems after multiple rounds of cut generation. To address this challenge, we propose a novel **dyn**amic **sep**arator configuration (**DynSep**) method that models separator configuration in different rounds as a reinforcement learning task, making decisions based on an incremental triplet graph updated by iteratively added cuts. Specifically, we tokenize the incremental subgraphs and utilize a decoder-only Transformer as our policy to autoregressively predict when to halt separation and which separators to activate at each round. Evaluated on synthetic and large-scale real-world MILP problems, DynSep speeds up average solving time by 64% on easy and medium datasets, and reduces primal-dual gap integral within the given time limit by 16% on hard datasets. Moreover, experiments demonstrate that DynSep well generalizes to MILP instances of significantly larger sizes than those seen during training.

In this work, we study the scalability of offline reinforcement learning (RL) algorithms. In principle, a truly scalable offline RL algorithm should be able to solve any given problem, regardless of its complexity, given sufficient data, compute, and model capacity. We investigate if and how current offline RL algorithms match up to this promise on diverse, challenging, previously unsolved tasks, using datasets up to 1000× larger than typical offline RL datasets. We observe that despite scaling up data, many existing offline RL algorithms exhibit poor scaling behavior, saturating well below the maximum performance. We hypothesize that the horizon is the main cause behind the poor scaling of offline RL. We empirically verify this hypothesis through several analysis experiments, showing that long horizons indeed present a fundamental barrier to scaling up offline RL. We then show that various horizon reduction techniques substantially enhance scalability on challenging tasks. Based on our insights, we also introduce a minimal yet scalable method named SHARSA that effectively reduces the horizon. SHARSA achieves the best asymptotic performance and scaling behavior among our evaluation methods, showing that explicitly reducing the horizon unlocks the scalability of offline RL.

Massively parallel GPU simulation environments have accelerated reinforcement learning (RL) research by enabling fast data collection for on-policy RL algorithms like Proximal Policy Optimization (PPO). To maximize throughput, it is common to use short rollouts per policy update, increasing the update-to-data (UTD) ratio. However, we find that, in this setting, standard synchronous resets introduce harmful nonstationarity, skewing the learning signal and destabilizing training. We introduce staggered resets, a simple yet effective technique where environments are initialized and reset at varied points within the task horizon. This yields training batches with more uniform state visitation distributions, reducing the nonstationarity induced by synchronized rollouts. We characterize dimensions along which RL environments can benefit significantly from staggered resets through toy environments. We then apply this technique to challenging high-dimensional robotics environments, achieving significantly higher sample efficiency, faster wall-clock convergence, and stronger final performance. Finally, this technique scales better with more parallel environments compared to naive synchronized rollouts, yielding more optimal utilization of computational resources.

The widespread deployment of Large Language Models (LLMs) trained on massive, uncurated corpora has raised growing concerns about the inclusion of sensitive, copyrighted, or illegal content. This has led to increasing interest in LLM unlearning: the task of selectively removing specific information from a model without retraining from scratch or degrading overall utility. However, existing methods often rely on large-scale forget and retain datasets, and suffer from unnatural responses, poor generalization, or catastrophic utility loss. In this work, we propose $\textbf{R}$einforcement $\textbf{U}$n$\textbf{LE}$arning ($\textbf{RULE}$), an efficient framework that formulates unlearning as a refusal boundary optimization problem. RULE is trained with a small portion of forget set and synthesized boundary queries, using a verifiable reward function that encourages safe refusal on forget-related queries while preserving helpful responses on permissible inputs. We provide both theoretical and empirical evidence demonstrating the effectiveness of RULE in achieving targeted unlearning without compromising model utility. Experimental results show that, with only 12\% forget set and 8\% synthesized boundary data, RULE outperforms existing baselines by up to $17.4\%$ forget quality and $16.3\%$ naturalness response while maintaining general utility, achieving $\textit{forget-retain Pareto Optimality}$. Remarkably, we further observe that RULE improves the $\textit{naturalness}$ of model outputs, enhances training $\textit{efficiency}$, and exhibits strong $\textit{generalization ability}$, generalizing refusal behavior to semantically related but unseen queries.

Bilevel reinforcement learning (BRL) has emerged as a powerful framework for aligning generative models, yet its theoretical foundations, especially sample complexity bounds, remain underexplored. In this work, we present the first sample complexity bound for BRL, establishing a rate of $\mathcal{O}(\epsilon^{-3})$ in continuous state-action spaces. Traditional MDP analysis techniques do not extend to BRL due to its nested structure and non-convex lower-level problems. We overcome these challenges by leveraging the Polyak-Łojasiewicz (PL) condition and the MDP structure to obtain closed-form gradients, enabling tight sample complexity analysis. Our analysis also extends to general bi-level optimization settings with non-convex lower levels, where we achieve state-of-the-art sample complexity results of $\mathcal{O}(\epsilon^{-3})$ improving upon existing bounds of $\mathcal{O}(\epsilon^{-6})$. Additionally, we address the computational bottleneck of hypergradient estimation by proposing a fully first-order, Hessian-free algorithm suitable for large-scale problems.

Comprehending natural language and following human instructions are critical capabilities for intelligent agents. However, the flexibility of linguistic instructions induces substantial ambiguity across language-conditioned tasks, severely degrading algorithmic performance. To address these limitations, we present a novel method named DAIL (Distributional Aligned Learning), featuring two key components: distributional policy and semantic alignment. Specifically, we provide theoretical results that the value distribution estimation mechanism enhances task differentiability. Meanwhile, the semantic alignment module captures the correspondence between trajectories and linguistic instructions. Extensive experimental results on both structured and visual observation benchmarks demonstrate that DAIL effectively resolves instruction ambiguities, achieving superior performance to baseline methods. Our implementation is available at https://github.com/RunpengXie/Distributional-Aligned-Learning.

Sparse-reward reinforcement learning (RL) can model a wide range of highly complex tasks. Solving sparse-reward tasks is RL's core premise — requiring efficient exploration coupled with long-horizon credit assignment — and overcoming these challenges is key for building self-improving agents with superhuman ability. We argue that solving complex and high-dimensional tasks requires solving simpler tasks that are *relevant* to the target task. In contrast, most prior work designs strategies for selecting exploratory tasks with the objective of solving *any* task, making exploration of challenging high-dimensional, long-horizon tasks intractable. We find that the sense of direction, necessary for effective exploration, can be extracted from existing reinforcement learning algorithms, without needing any prior information. Based on this finding, we propose a method for _**di**rected **s**parse-reward goal-**co**nditioned **ve**ry long-horizon **R**L_ (DISCOVER), which selects exploratory goals in the direction of the target task. We connect DISCOVER to principled exploration in bandits, formally bounding the time until the target task becomes achievable in terms of the agent's initial distance to the target, but independent of the volume of the space of all tasks. Empirically, we perform a thorough evaluation in high-dimensional simulated environments. We find that the directed goal selection of DISCOVER solves exploration problems that are beyond the reach of prior state-of-the-art exploration methods in RL.

We consider how to construct state abstractions compatible with a given set of abstract actions, to obtain a well-formed abstract Markov decision process (MDP). We show that the Bellman equation suggests that abstract states should represent distributions over states in the ground MDP; we characterize the conditions under which the resulting process is Markov and approximately model-preserving, derive algorithms for constructing and planning with the abstract MDP, and apply them to a visual maze task. We generalize these results to the factored actions case, characterizing the conditions that result in factored abstract states and apply the resulting algorithm to Montezuma's Revenge. These results provide a powerful and principled framework for constructing neurosymbolic abstract Markov decision processes.

Offline goal-conditioned reinforcement learning (GCRL) offers a practical learning paradigm in which goal-reaching policies are trained from abundant state–action trajectory datasets without additional environment interaction. However, offline GCRL still struggles with long-horizon tasks, even with recent advances that employ hierarchical policy structures, such as HIQL. Identifying the root cause of this challenge, we observe the following insight. Firstly, performance bottlenecks mainly stem from the high-level policy’s inability to generate appropriate subgoals. Secondly, when learning the high-level policy in the long-horizon regime, the sign of the advantage estimate frequently becomes incorrect. Thus, we argue that improving the value function to produce a clear advantage estimate for learning the high-level policy is essential. In this paper, we propose a simple yet effective solution: _**Option-aware Temporally Abstracted**_ value learning, dubbed **OTA**, which incorporates temporal abstraction into the temporal-difference learning process. By modifying the value update to be _option-aware_, our approach contracts the effective horizon length, enabling better advantage estimates even in long-horizon regimes. We experimentally show that the high-level policy learned using the OTA value function achieves strong performance on complex tasks from OGBench, a recently proposed offline GCRL benchmark, including maze navigation and visual robotic manipulation environments. Our code is available at https://github.com/ota-v/ota-v

Continual learning is often motivated by the idea, known as the big world hypothesis, that the "world is bigger" than the agent. Recent problem formulations capture this idea by explicitly constraining an agent relative to the environment. These constraints lead to solutions in which the agent continually adapts to best use its limited capacity, rather than converging to a fixed solution. However, explicit constraints can be ad hoc, difficult to incorporate, and limiting to the effectiveness of scaling up the agent's capacity. In this paper, we characterize a problem setting in which an agent, regardless of its capacity, is implicitly constrained by being embedded in the environment. In particular, we introduce a computationally-embedded perspective that represents an embedded agent as an automaton simulated within a universal (formal) computer. We prove that such an automaton is implicitly constrained and that it is equivalent to an agent that interacts with a partially observable Markov decision process over a countably infinite state-space. We then propose an objective for this setting, which we call interactivity, that measures an agent's ability to continually adapt its behaviour and to continually learn new predictions. We develop a reinforcement learning algorithm for maximizing interactivity and a synthetic benchmark to experimentation on continual learning. Our results indicate that deep nonlinear networks struggle to sustain interactivity whereas deep linear networks can achieve higher interactivity as capacity increases.

Direct Preference Optimization (DPO) is a powerful paradigm to align language models with human preferences using pairwise comparisons. However, its binary win-or-loss supervision often proves insufficient for training small models with limited capacity. Prior works attempt to distill information from large teacher models using behavior cloning or KL divergence. These methods often focus on mimicking current behavior and overlook distilling reward modeling. To address this issue, we propose \textit{Teacher Value-based Knowledge Distillation} (TVKD), which introduces an auxiliary reward from the value function of the teacher model to provide a soft guide. This auxiliary reward is formulated to satisfy potential-based reward shaping, ensuring that the global reward structure and optimal policy of DPO are preserved. TVKD can be integrated into the standard DPO training framework and does not require additional rollouts. Our experimental results show that TVKD consistently improves performance across various benchmarks and model sizes.

Recent work on large language models has demonstrated the use of model-free reinforcement learning (RL) to train reasoning-like capabilities. The emergence of "thinking" through model-free RL is interesting as thinking actions neither produce reward nor change the external world state to one where the agent is more likely to get reward. This paper seeks to build a domain-independent understanding of when model-free RL will lead to such "thinking" as a strategy for reward maximization. To build this understanding, we first introduce a theoretical model which we call a thought Markov decision process (MDP). Thought MDPs minimally extend the classical MDP model to include an abstract notion of thought state and thought action. Using the thought MDP model, we prove the importance of policy initialization in determining whether or not thinking emerges and show formally that thought actions are equivalent to the agent choosing to perform a step of policy improvement before continuing to act. We then show that open-source LLMs satisfy the conditions that our theory predicts are necessary for model-free RL to produce thinking-like behavior. Finally, we hypothesize sufficient conditions that would enable thinking to be learned outside of language generation and introduce a toy domain where a combination of multi-task pre-training and designated thought actions enable more data-efficient RL compared to non-thinking agents.