NeurIPS 2025 Papers

This paper studies reinforcement learning from human feedback (RLHF) for aligning large language models with human preferences. While RLHF has demonstrated promising results, many algorithms are highly sensitive to misspecifications in the underlying preference model (e.g., the Bradley-Terry model), the reference policy, or the reward function, resulting in undesirable fine-tuning. To address model misspecification, we propose a doubly robust preference optimization algorithm that remains consistent when either the preference model or the reference policy is correctly specified (without requiring both). Our proposal demonstrates superior and more robust performance than state-of-the-art algorithms, both in theory and in practice. The code is available at https://github.com/DRPO4LLM/DRPO4LLM

Reinforcement learning (RL) is increasingly applied to real-world problems involving complex and structured decisions, such as routing, scheduling, and assortment planning. These settings challenge standard RL algorithms, which struggle to scale, generalize, and exploit structure in the presence of combinatorial action spaces. We propose Structured Reinforcement Learning (SRL), a novel actor-critic paradigm that embeds combinatorial optimization-layers into the actor neural network. We enable end-to-end learning of the actor via Fenchel-Young losses and provide a geometric interpretation of SRL as a primal-dual algorithm in the dual of the moment polytope. Across six environments with exogenous and endogenous uncertainty, SRL matches or surpasses the performance of unstructured RL and imitation learning on static tasks and improves over these baselines by up to 92\% on dynamic problems, with improved stability and convergence speed.

Deep reinforcement learning (RL) agents frequently suffer from neuronal activity loss, which impairs their ability to adapt to new data and learn continually. A common method to quantify and address this issue is the $\tau$-dormant neuron ratio, which uses activation statistics to measure the expressive ability of neurons. While effective for simple MLP-based agents, this approach loses statistical power in more complex architectures. To address this, we argue that in advanced RL agents, maintaining a neuron's **learning capacity**, its ability to adapt via gradient updates, is more critical than preserving its expressive ability. Based on this insight, we shift the statistical objective from activations to gradients, and introduce **GraMa** (**Gra**dient **Ma**gnitude Neural Activity Metric), a lightweight, architecture-agnostic metric for quantifying neuron-level learning capacity. We show that **GraMa** effectively reveals persistent neuron inactivity across diverse architectures, including residual networks, diffusion models, and agents with varied activation functions. Moreover, **re**setting neurons guided by **GraMa** (**ReGraMa**) consistently improves learning performance across multiple deep RL algorithms and benchmarks, such as MuJoCo and the DeepMind Control Suite. **We make our code available.**

We present DEPS, an end-to-end algorithm for discovering parameterized skills from expert demonstrations. Our method learns parameterized skill policies jointly with a meta-policy that selects the appropriate discrete skill and continuous parameters at each timestep. Using a combination of temporal variational inference and information-theoretic regularization methods, we address the challenge of degeneracy common in latent variable models, ensuring that the learned skills are temporally extended, semantically meaningful, and adaptable. We empirically show that learning parameterized skills from multitask expert demonstrations significantly improves generalization to unseen tasks. Our method outperforms multitask as well as skill learning baselines on both LIBERO and MetaWorld benchmarks. We also demonstrate that DEPS discovers interpretable parameterized skills, such as an object grasping skill whose continuous arguments define the grasp location.

Preference-based reinforcement learning (PbRL) bypasses complex reward engineering by learning from human feedback. However, due to the high cost of obtaining feedback, PbRL typically relies on a limited set of preference-labeled samples. This data scarcity introduces two key inefficiencies: (1) the reward model overfits to the limited feedback, leading to poor generalization to unseen samples, and (2) the agent exploits the learned reward model, exacerbating overestimation of action values in temporal difference (TD) learning. To address these issues, we propose STAR, an efficient PbRL method that integrates preference margin regularization and policy regularization. Preference margin regularization mitigates overfitting by introducing a bounded margin in reward optimization, preventing excessive bias toward specific feedback. Policy regularization bootstraps a conservative estimate $\widehat{Q}$ from well-supported state-action pairs in the replay memory, reducing overestimation during policy learning. Experimental results show that STAR improves feedback efficiency, achieving 34.8\% higher performance in online settings and 29.7\% in offline settings compared to state-of-the-art methods. Ablation studies confirm that STAR facilitates more robust reward and value function learning. The videos of this project are released at https://sites.google.com/view/pbrl-star.

Traditional reinforcement learning (RL) assumes the agents make decisions based on Markov decision processes (MDPs) with one-step transition models. In many real-world applications, such as energy management and stock investment, agents can access multi-step predictions of future states, which provide additional advantages for decision making. However, multi-step predictions are inherently high-dimensional: naively embedding these predictions into an MDP leads to an exponential blow-up in state space and the curse of dimensionality. Moreover, existing RL theory provides few tools to analyze prediction-augmented MDPs, as it typically works on one-step transition kernels and cannot accommodate multi-step predictions with errors or partial action-coverage. We address these challenges with three key innovations: First, we propose the \emph{Bayesian value function} to characterize the optimal prediction-aware policy tractably. Second, we develop a novel \emph{Bellman–Jensen Gap} analysis on the Bayesian value function, which enables characterizing the value of imperfect predictions. Third, we introduce BOLA (Bayesian Offline Learning with Online Adaptation), a two-stage model-based RL algorithm that separates offline Bayesian value learning from lightweight online adaptation to real-time predictions. We prove that BOLA remains sample-efficient even under imperfect predictions. We validate our theory and algorithm on synthetic MDPs and a real-world wind energy storage control problem.

Real-world robotic tasks often require agents to achieve sequences of goals while respecting time-varying safety constraints. However, standard Reinforcement Learning (RL) paradigms are fundamentally limited in these settings. A natural approach to these problems is to combine RL with Linear-time Temporal Logic (LTL), a formal language for specifying complex, temporally extended tasks and safety constraints. Yet, existing RL methods for LTL objectives exhibit poor empirical performance in complex and continuous environments. As a result, no scalable methods support both temporally ordered goals and safety simultaneously, making them ill-suited for realistic robotics scenarios. We propose Automaton Constrained Q-Learning (ACQL), an algorithm that addresses this gap by combining goal-conditioned value learning with automaton-guided reinforcement. ACQL supports most LTL task specifications and leverages their automaton representation to explicitly encode stage-wise goal progression and both stationary and non-stationary safety constraints. We show that ACQL outperforms existing methods across a range of continuous control tasks, including cases where prior methods fail to satisfy either goal-reaching or safety constraints. We further validate its real-world applicability by deploying ACQL on a 6-DOF robotic arm performing a goal-reaching task in a cluttered, cabinet-like space with safety constraints. Our results demonstrate that ACQL is a robust and scalable solution for learning robotic behaviors according to rich temporal specifications.

Many real-world control problems require continual policy adjustments to balance multiple objectives, which requires the acquisition of high-quality policies to cover diverse preferences. Multi-Objective Reinforcement Learning (MORL) provides a general framework to solve such problems. However, current MORL methods suffer from high sample complexity, primarily due to the neglect of efficient knowledge sharing and conflicts in optimization with different preferences. To this end, this paper introduces a novel framework, Conflict Objective Regularization in Latent Space (**COLA**). To enable efficient knowledge sharing, COLA establishes a shared latent representation space for common knowledge, which can avoid redundant learning under different preferences. Besides, COLA introduces a regularization term for the value function to mitigate the negative effects of conflicting preferences on the value function approximation, thereby improving the accuracy of value estimation. The experimental results across various multi-objective continuous control tasks demonstrate the significant superiority of COLA over the state-of-the-art MORL baselines. Code is available at https://github.com/yeshenpy/COLA.

Inspired by the success of DeepSeek-R1, we explore the potential of rule-based reinforcement learning (RL) in MLLM post-training for perception policy learning. While promising, our initial experiments reveal that incorporating a thinking process through RL does not consistently lead to performance gains across all visual perception tasks. This leads us to delve into the essential role of RL in the context of visual perception. In this work, we return to the fundamentals and explore the effects of RL on different perception tasks. We observe that the perceptual perplexity is a major factor in determining the effectiveness of RL. We also observe that reward design plays a crucial role in further approaching the upper limit of model perception. To leverage these findings, we propose Perception-R1, a scalable RL framework using GRPO during MLLM post-training. With a standard Qwen2-VL-2B-Instruct, Perception-R1 achieves +4.2% on RefCOCO+, +17.9% on PixMo-Count, +4.2% on PageOCR, and notably, 31.9% AP on COCO2017 val for the first time, establishing a strong baseline for perception policy learning.

Risk-averse total-reward Markov Decision Processes (MDPs) offer a promising framework for modeling and solving undiscounted infinite-horizon objectives. Existing model-based algorithms for risk measures like the entropic risk measure (ERM) and entropic value-at-risk (EVaR) are effective in small problems, but require full access to transition probabilities. We propose a Q-learning algorithm to compute the optimal stationary policy for total-reward ERM and EVaR objectives with strong convergence and performance guarantees. The algorithm and its optimality are made possible by ERM's dynamic consistency and elicitability. Our numerical results on tabular domains demonstrate quick and reliable convergence of the proposed Q-learning algorithm to the optimal risk-averse value function.

In the zero-shot policy transfer setting in reinforcement learning, the goal is to train an agent on a fixed set of training environments so that it can generalise to similar, but unseen, testing environments. Previous work has shown that policy distillation after training can sometimes produce a policy that outperforms the original in the testing environments. However, it is not yet entirely clear why that is, or what data should be used to distil the policy. In this paper, we prove, under certain assumptions, a generalisation bound for policy distillation after training. The theory provides two practical insights: for improved generalisation, you should 1) train an ensemble of distilled policies, and 2) distil it on as much data from the training environments as possible. We empirically verify that these insights hold in more general settings, when the assumptions required for the theory no longer hold. Finally, we demonstrate that an ensemble of policies distilled on a diverse dataset can generalise significantly better than the original agent.

Motivated by practical applications where stable long-term performance is critical—such as robotics, operations research, and healthcare—we study the problem of distributionally robust (DR) average-reward reinforcement learning. We propose two algorithms that achieve near-optimal sample complexity. The first reduces the problem to a DR discounted Markov decision process (MDP), while the second, Anchored DR Average-Reward MDP, introduces an anchoring state to stabilize the controlled transition kernels within the uncertainty set. Assuming the nominal MDP is uniformly ergodic, we prove that both algorithms attain a sample complexity of $\widetilde{O}\left(|\mathbf{S}||\mathbf{A}| t_{\mathrm{mix}}^2\varepsilon^{-2}\right)$ for estimating the optimal policy as well as the robust average reward under KL and $f_k$-divergence-based uncertainty sets, provided the uncertainty radius is sufficiently small. Here, $\varepsilon$ is the target accuracy, $|\mathbf{S}|$ and $|\mathbf{A}|$ denote the sizes of the state and action spaces, and $t_{\mathrm{mix}}$ is the mixing time of the nominal MDP. This represents the first finite-sample convergence guarantee for DR average-reward reinforcement learning. We further validate the convergence rates of our algorithms through numerical experiments.

Reinforcement learning (RL) heavily depends on well-designed reward functions, which are often biased and difficult to design for complex behaviors. Preference-based RL (PbRL) addresses this by learning reward models from human feedback, but its practicality is constrained by a critical dilemma: while existing methods reduce human effort through query optimization, they neglect the preference buffer's restricted coverage — a factor that fundamentally determines the reliability of reward model. We systematically demonstrate this limitation creates distributional mismatch: reward models trained on static buffers reliably assess in-distribution trajectories but falter with out-of-distribution (OOD) trajectories from policy exploration. Crucially, such failures in policy-proximal regions directly misguide iterative policy updates. To address this, we propose **Proximal Policy Exploration (PPE)** with two key components: (1) a *proximal-policy extension* method that expands exploration in undersampled policy-proximal regions, and (2) a *mixture distribution query* method that balances in-distribution and OOD trajectory sampling. By enhancing buffer coverage while preserving evaluation accuracy in policy-proximal regions, PPE enables more reliable policy updates. Experiments across continuous control tasks demonstrate that PPE enhances preference feedback utilization efficiency and RL sample efficiency over baselines, highlighting preference buffer coverage management's vital role in PbRL.

This paper investigates infinite-horizon average reward Constrained Markov Decision Processes (CMDPs) under general parametrized policies with smooth and bounded policy gradients. We propose a Primal-Dual Natural Actor-Critic algorithm that adeptly manages constraints while ensuring a high convergence rate. In particular, our algorithm achieves global convergence and constraint violation rates of $\tilde{\mathcal{O}}(1/\sqrt{T})$ over a horizon of length $T$ when the mixing time, $\tau_{\mathrm{mix}}$, is known to the learner. In absence of knowledge of $\tau_{\mathrm{mix}}$, the achievable rates change to $\tilde{\mathcal{O}}(1/T^{0.5-\epsilon})$ provided that $T \geq \tilde{\mathcal{O}}\left(\tau_{\mathrm{mix}}^{2/\epsilon}\right)$. Our results match the theoretical lower bound for Markov Decision Processes and establish a new benchmark in the theoretical exploration of average reward CMDPs.

In recent years, the successor representation (SR) has attracted increasing attention in reinforcement learning (RL), and it has been used to address some of its key challenges, such as exploration, credit assignment, and generalization. The SR can be seen as representing the underlying credit assignment structure of the environment by implicitly encoding its induced transition dynamics. However, the SR is reward-agnostic. In this paper, we discuss a similar representation that also takes into account the reward dynamics of the problem. We study the default representation (DR), a recently proposed representation with limited theoretical (and empirical) analysis. Here, we lay some of the theoretical foundation underlying the DR in the tabular case by (1) deriving dynamic programming and (2) temporal-difference methods to learn the DR, (3) characterizing the basis for the vector space of the DR, and (4) formally extending the DR to the function approximation case through default features. Empirically, we analyze the benefits of the DR in many of the settings in which the SR has been applied, including (1) reward shaping, (2) option discovery, (3) exploration, and (4) transfer learning. Our results show that, compared to the SR, the DR gives rise to qualitatively different, reward-aware behaviour and quantitatively better performance in several settings.

Progress in offline reinforcement learning (RL) has been impeded by ambiguous problem definitions and entangled algorithmic designs, resulting in inconsistent implementations, insufficient ablations, and unfair evaluations. Although offline RL explicitly avoids environment interaction, prior methods frequently employ extensive, undocumented online evaluation for hyperparameter tuning, complicating method comparisons. Moreover, existing reference implementations differ significantly in boilerplate code, obscuring their core algorithmic contributions. We address these challenges by first introducing a rigorous taxonomy and a transparent evaluation protocol that explicitly quantifies online tuning budgets. To resolve opaque algorithmic design, we provide clean, minimalistic, single-file implementations of various model-free and model-based offline RL methods, significantly enhancing clarity and achieving substantial speed-ups. Leveraging these streamlined implementations, we propose Unifloral, a unified algorithm that encapsulates diverse prior approaches and enables development within a single, comprehensive hyperparameter space. Using Unifloral with our rigorous evaluation protocol, we develop two novel algorithms - TD3-AWR (model-free) and MoBRAC (model-based) - which substantially outperform established baselines. Our implementation is publicly available at https://github.com/EmptyJackson/unifloral.

Direct Preference Optimization (DPO) has emerged as an effective approach for aligning large language models (LLMs) with human preferences. However, its performance is highly dependent on the quality of the underlying human preference data. To address this bottleneck, prior work has explored various data selection strategies, but these methods often overlook the impact of the evolving states of the language model during the optimization process. In this paper, we introduce a novel problem: Sample Scheduling for DPO, which aims to dynamically and adaptively schedule training samples based on the model's evolving batch-wise states throughout preference optimization. To solve this problem, we propose SamS, an efficient and effective algorithm that adaptively selects samples in each training batch based on the LLM's learning feedback to maximize the potential generalization performance. Notably, without modifying the core DPO algorithm, simply integrating SamS significantly improves performance across tasks, with minimal additional computational overhead. This work points to a promising new direction for improving LLM alignment through batch-wise sample selection, with potential generalization to RLHF and broader supervised learning paradigms.

Offline reinforcement learning (RL) aims to learn a policy from a fixed dataset without additional environment interaction. However, effective offline policy learning often requires a large and diverse dataset to mitigate epistemic uncertainty. Collecting such data demands substantial online interactions, which are costly or infeasible in many real-world domains. Therefore, improving policy learning from limited offline data—achieving high data efficiency—is critical for practical offline RL. In this paper, we propose a simple yet effective plug-and-play pretraining framework that initializes the feature representation of a $Q$-network to enhance data efficiency in offline RL. Our approach employs a shared $Q$-network architecture trained in two stages: pretraining a backbone feature extractor with a transition prediction head; training a $Q$-network—combining the backbone feature extractor and a $Q$-value head—with *any* offline RL objective. Extensive experiments on the D4RL, Robomimic, V-D4RL, and ExoRL benchmarks show that our method substantially improves both performance and data efficiency across diverse datasets and domains. Remarkably, with only **10\%** of the dataset, our approach outperforms standard offline RL baselines trained on the full data.

Incorporating pre-collected offline data from a source environment can significantly improve the sample efficiency of reinforcement learning (RL), but this benefit is often challenged by discrepancies between the transition dynamics of the source and target environments. Existing methods typically address this issue by penalizing or filtering out source transitions in high dynamics-gap regions. However, their estimation of the dynamics gap often relies on KL divergence or mutual information, which can be ill-defined when the source and target dynamics have disjoint support. To overcome these limitations, we propose CompFlow, a method grounded in the theoretical connection between flow matching and optimal transport. Specifically, we model the target dynamics as a conditional flow built upon the output distribution of the source-domain flow, rather than learning it directly from a Gaussian prior. This composite structure offers two key advantages: (1) improved generalization for learning target dynamics, and (2) a principled estimation of the dynamics gap via the Wasserstein distance between source and target transitions. Leveraging our principled estimation of the dynamics gap, we further introduce an optimistic active data collection strategy that prioritizes exploration in regions of high dynamics gap, and theoretically prove that it reduces the performance disparity with the optimal policy. Empirically, CompFlow outperforms strong baselines across several RL benchmarks with shifted dynamics.