ICML 2025 Papers

Multi-agent reinforcement learning (MARL) has achieved significant progress in large-scale traffic control, autonomous vehicles, and robotics. Drawing inspiration from biological systems where roles naturally emerge to enable coordination, role-based MARL methods have been proposed to enhance cooperation learning for complex tasks. However, existing methods exclusively derive roles from an agent's past experience during training, neglecting their influence on its future trajectories. This paper introduces a key insight: an agent’s role should shape its future behavior to enable effective coordination. Hence, we propose Role Discovery and Diversity through Dynamics Models (R3DM), a novel role-based MARL framework that learns emergent roles by maximizing the mutual information between agents' roles, observed trajectories, and expected future behaviors. R3DM optimizes the proposed objective through contrastive learning on past trajectories to first derive intermediate roles that shape intrinsic rewards to promote diversity in future behaviors across different roles through a learned dynamics model. Benchmarking on SMAC and SMACv2 environments demonstrates that R3DM outperforms state-of-the-art MARL approaches, improving multi-agent coordination to increase win rates by up to 20%. The code is available at https://github.com/UTAustin-SwarmLab/R3DM.

Reinforcement learning from human feedback (RLHF), which aligns a diffusion model with input prompt, has become a crucial step in building reliable generative AI models. Most works in this area uses a *discrete-time* formulation, which is prone to induced errors, and often not applicable to models with higher-order/black-box solvers. The objective of this study is to develop a disciplined approach to fine-tuning diffusion models using *continuous-time* RL, formulated as a stochastic control problem with a reward function that aligns the end result (terminal state) with input prompt. The key idea is to treat score matching as controls or actions, and thereby connecting to policy optimization and regularization in continuous-time RL. To carry out this idea, we lay out a new policy optimization framework for continuous-time RL, and illustrate its potential in enhancing the value networks design space via leveraging the structural property of diffusion models. We validate the advantages of our method by experiments in downstream tasks of fine-tuning large-scale Text2Image models, Stable Diffusion v1.5.

Rainbow Deep Q-Network (DQN) demonstrated combining multiple independent enhancements could significantly boost a reinforcement learning (RL) agent’s performance. In this paper, we present “Beyond The Rainbow” (BTR), a novel algorithm that integrates six improvements from across the RL literature to Rainbow DQN, establishing a new state-of-the-art for RL using a desktop PC, with a human-normalized interquartile mean (IQM) of 7.6 on Atari-60. Beyond Atari, we demonstrate BTR’s capability to handle complex 3D games, successfully training agents to play Super Mario Galaxy, Mario Kart, and Mortal Kombat with minimal algorithmic changes. Designing BTR with computational efficiency in mind, agents can be trained using a high-end desktop PC on 200 million Atari frames within 12 hours. Additionally, we conduct detailed ablation studies of each component, analyzing the performance and impact using numerous measures.

Multi-objective Reinforcement Learning (MORL) seeks to develop policies that simultaneously optimize multiple conflicting objectives, but it requires extensive online interactions. Offline MORL provides a promising solution by training on pre-collected datasets to generalize to any preference upon deployment. However, real-world offline datasets are often conservatively and narrowly distributed, failing to comprehensively cover preferences, leading to the emergence of out-of-distribution (OOD) preference areas. Existing offline MORL algorithms exhibit poor generalization to OOD preferences, resulting in policies that do not align with preferences. Leveraging the excellent expressive and generalization capabilities of diffusion models, we propose MODULI (Multi-objective Diffusion Planner with Sliding Guidance), which employs a preference-conditioned diffusion model as a planner to generate trajectories that align with various preferences and derive action for decision-making. To achieve accurate generation, MODULI introduces two return normalization methods under diverse preferences for refining guidance. To further enhance generalization to OOD preferences, MODULI proposes a novel sliding guidance mechanism, which involves training an additional slider adapter to capture the direction of preference changes. Incorporating the slider, it transitions from in-distribution (ID) preferences to generating OOD preferences, patching, and extending the incomplete Pareto front. Extensive experiments on the D4MORL benchmark demonstrate that our algorithm outperforms state-of-the-art Offline MORL baselines, exhibiting excellent generalization to OOD preferences.

Modeling human preferences is crucial for aligning foundation models with human values. Traditional reward modeling methods, such as the Bradley-Terry (BT) reward model, fall short in expressiveness, particularly in addressing intransitive preferences. In this paper, we introduce \emph{preference embedding}, an approach that embeds responses into a latent space to capture intricate preference structures efficiently, achieving linear query complexity. Additionally, we propose preference score-based General Preference Optimization (GPO), which generalizes reward-based reinforcement learning from human feedback (RLHF). Experimental results show that our General Preference embedding Model (GPM) consistently outperforms the BT reward model on the RewardBench benchmark and effectively models cyclic preferences where any BT reward model behaves like a random guess. Furthermore, evaluations on downstream tasks such as AlpacaEval2.0, following the language model post-training with GPO and our general preference model, reveal performance improvements over BT models. These findings indicate that our method may enhance the alignment of foundation models with nuanced human values. The code is available at https://github.com/general-preference/general-preference-model.

Teaching large language models (LLMs) to critique and refine their outputs is crucial for building systems that can iteratively improve, yet it is fundamentally limited by the ability to provide *accurate judgments* and *actionable suggestions*. In this work, we study LLM critics for code generation and propose $\texttt{CTRL}$, a framework for $\texttt{C}$ritic $\texttt{T}$raining via $\texttt{R}$einforcement $\texttt{L}$earning, which trains a critic model to generate feedback that maximizes correction performance for a fixed generator model without human supervision. Our results demonstrate that critics trained with $\texttt{CTRL}$ significantly enhance pass rates and mitigate compounding errors across both base and stronger generator models. Furthermore, we show that these critic models act as accurate generative reward models and enable test-time scaling through iterative critique-revision, achieving up to 106.1\% relative improvements across challenging code generation benchmarks.

Foundation Models (FMs) and World Models (WMs) offer complementary strengths in task generalization at different levels. In this work, we propose FOUNDER, a framework that integrates the generalizable knowledge embedded in FMs with the dynamic modeling capabilities of WMs to enable open-ended task solving in embodied environments in a reward-free manner. We learn a mapping function that grounds FM representations in the WM state space, effectively inferring the agent's physical states in the world simulator from external observations. This mapping enables the learning of a goal-conditioned policy through imagination during behavior learning, with the mapped task serving as the goal state. Our method leverages the predicted temporal distance to the goal state as an informative reward signal. FOUNDER demonstrates superior performance on various multi-task offline visual control benchmarks, excelling in capturing the deep-level semantics of tasks specified by text or videos, particularly in scenarios involving complex observations or domain gaps where prior methods struggle. The consistency of our learned reward function with the ground-truth reward is also empirically validated. Our project website is https://sites.google.com/view/founder-rl.

We present Gradient Boosting Reinforcement Learning (GBRL), a framework that adapts the strengths of Gradient Boosting Trees (GBT) to reinforcement learning (RL) tasks. While neural networks (NNs) have become the de facto choice for RL, they face significant challenges with structured and categorical features and tend to generalize poorly to out-of-distribution samples. These are challenges for which GBTs have traditionally excelled in supervised learning. However, GBT's application in RL has been limited. The design of traditional GBT libraries is optimized for static datasets with fixed labels, making them incompatible with RL's dynamic nature, where both state distributions and reward signals evolve during training. GBRL overcomes this limitation by continuously interleaving tree construction with environment interaction. Through extensive experiments, we demonstrate that GBRL outperforms NNs in domains with structured observations and categorical features, while maintaining competitive performance on standard continuous control benchmarks. Like its supervised learning counterpart, GBRL demonstrates superior robustness to out-of-distribution samples and better handles irregular state-action relationships.

In the standard Reinforcement Learning (RL) paradigm, the action space is assumed to be fixed and immutable throughout the learning process. However, in many real-world scenarios, not all actions are available at every decision stage. The available action set may depend on the current environment state, domain-specific constraints, or other (potentially stochastic) factors outside the agent's control. To address these realistic scenarios, we introduce a novel paradigm called *Sleeping Reinforcement Learning*, where the available action set varies during the interaction with the environment. We start with the simpler scenario in which the available action sets are revealed at the beginning of each episode. We show that a modification of UCBVI achieves regret of order $\widetilde{\mathcal{O}}(H\sqrt{SAT})$, where $H$ is the horizon, $S$ and $A$ are the cardinalities of the state and action spaces, respectively, and $T$ is the learning horizon. Next, we address the more challenging and realistic scenario in which the available actions are disclosed only at each decision stage. By leveraging a novel construction, we establish a minimax lower bound of order $\Omega(\sqrt{T 2^{A/2}})$ when the availability of actions is governed by a Markovian process, establishing a statistical barrier of the problem. Focusing on the statistically tractable case where action availability depends only on the current state and stage, we propose a new optimistic algorithm that achieves regret guarantees of order $\widetilde{\mathcal{O}}(H\sqrt{SAT})$, showing that the problem shares the same complexity of standard RL.

Multi-task reinforcement learning (RL) encounters significant challenges due to varying task complexities and their reward distributions from the environment. To address these issues, in this paper, we propose Adaptive Reward Scaling (ARS), a novel framework that dynamically adjusts reward magnitudes and leverages a periodic network reset mechanism. ARS introduces a history-based reward scaling strategy that ensures balanced reward distributions across tasks, enabling stable and efficient training. The reset mechanism complements this approach by mitigating overfitting and ensuring robust convergence. Empirical evaluations on the Meta-World benchmark demonstrate that ARS significantly outperforms baseline methods, achieving superior performance on challenging tasks while maintaining overall learning efficiency. These results validate ARS's effectiveness in tackling diverse multi-task RL problems, paving the way for scalable solutions in complex real-world applications.

Training models to efficiently use test-time compute is crucial for improving the reasoning performance of LLMs. While current methods mostly do so via fine-tuning on search traces or running RL against the 0/1 outcome reward, do these approaches efficiently utilize test-time compute? Would these approaches continue to scale as the budget improves? In this paper, we try to answer these questions. We formalize the problem of optimizing test-time compute as a meta reinforcement learning (RL) problem, which provides a principled perspective on spending test-time compute from the lens of exploration and exploitation. It also motivates the use of cumulative regret to measure the efficacy of test-time compute by viewing a long output stream as consisting of several episodes from the model. While current state-of-the-art models do not optimize regret, we show that regret can be minimized by running final 0/1 reward RL regularized by a dense reward bonus, given by the "information gain" from each subsequent block in the output stream. We prescribe an approach for quantifying information gain, which measures the utility of an intermediate segment of tokens towards improving accuracy of the final answer. We instantiate this idea to develop MRT, a new class of finetuning methods for optimizing test-time compute. Fine-tuning with MRT leads to substantial improvements in both performance and token efficiency on the AIME dataset.

The Bradley-Terry (BT) model is widely practiced in reward modeling for reinforcement learning with human feedback (RLHF). Despite its effectiveness, reward models (RMs) trained with BT model loss as one-way classifiers are prone to over-optimization, losing generalizability to unseen inputs. In this paper, we study the cause of over-optimization and its downstream effects on the RLHF procedure, highlighting the importance of robustness in RMs. First, we show that the excessive dispersion of hidden state norms is the main source of over-optimization. Correspondingly, we propose batch-wise sum-to-zero regularization (BSR) that enforces reward sum for each batch to be zero-centered, constraining the rewards with abnormally large magnitudes. We assess the impact of BSR in improving robustness in RMs through four scenarios of over-optimization, where BSR consistently manifests better robustness on unseen inputs. Then, we compare the plain BT model and BSR on RLHF training and empirically show that robust RMs better align the policy to the gold preference model. Finally, we apply BSR to high-quality data and models, which surpasses state-of-the-art RMs in the 8B scale by adding more than 5\% in complex preference prediction tasks. By conducting RLOO training with 8B RMs, AlpacaEval 2.0, with reducing generation length by 40\% while adding a 7\% increase in win rate, further highlights that robustness in RMs induces robustness in RLHF training.

Distributionally robust policy learning aims to find a policy that performs well under the worst-case distributional shift, and yet most existing methods for robust policy learning consider the worst-case *joint* distribution of the covariate and the outcome. The joint-modeling strategy can be unnecessarily conservative when we have more information on the source of distributional shifts. This paper studies a more nuanced problem --- robust policy learning under the *concept drift*, when only the conditional relationship between the outcome and the covariate changes. To this end, we first provide a doubly-robust estimator for evaluating the worst-case average reward of a given policy under a set of perturbed conditional distributions. We show that the policy value estimator enjoys asymptotic normality even if the nuisance parameters are estimated with a slower-than-root-$n$ rate. We then propose a learning algorithm that outputs the policy maximizing the estimated policy value within a given policy class $\Pi$, and show that the sub-optimality gap of the proposed algorithm is of the order $\kappa(\Pi)n^{-1/2}$, where $\kappa(\Pi)$ is the entropy integral of $\Pi$ under the Hamming distance and $n$ is the sample size. A matching lower bound is provided to show the optimality of the rate. The proposed methods are implemented and evaluated in numerical studies, demonstrating substantial improvement compared with existing benchmarks.

Recent advancements in reinforcement learning from human feedback have shown that utilizing fine-grained token-level reward models can substantially enhance the performance of Proximal Policy Optimization (PPO) in aligning large language models. However, it is challenging to leverage such token-level reward as guidance for Direct Preference Optimization (DPO), since DPO is formulated as a sequence-level bandit problem. To address this challenge, this work decomposes the sequence-level PPO into a sequence of token-level proximal policy optimization problems and then frames the problem of token-level PPO with token-level reward guidance, from which closed-form optimal token-level policy and the corresponding token-level reward can be derived. Using the obtained reward and Bradley-Terry model, this work establishes a framework of computable loss functions with token-level reward guidance for DPO, and proposes a practical reward guidance based on the induced DPO reward. This formulation enables different tokens to exhibit varying degrees of deviation from reference policy based on their respective rewards. Experiment results demonstrate that our method achieves substantial performance improvements over DPO, with win rate gains of up to 7.5 points on MT-Bench, 6.2 points on AlpacaEval 2, and 4.3 points on Arena-Hard. Code is available at https://github.com/dvlab-research/TGDPO.

Learning to cooperate in distributed partially observable environments with no communication abilities poses significant challenges for multi-agent deep reinforcement learning (MARL). This paper addresses key concerns in this domain, focusing on inferring state representations from individual agent observations and leveraging these representations to enhance agents' exploration and collaborative task execution policies. To this end, we propose a novel state modelling framework for cooperative MARL, where agents infer meaningful belief representations of the non-observable state, with respect to optimizing their own policies, while filtering redundant and less informative joint state information. Building upon this framework, we propose the MARL SMPE$^2$ algorithm. In SMPE$^2$, agents enhance their own policy's discriminative abilities under partial observability, explicitly by incorporating their beliefs into the policy network, and implicitly by adopting an adversarial type of exploration policies which encourages agents to discover novel, high-value states while improving the discriminative abilities of others. Experimentally, we show that SMPE$^2$ outperforms a plethora of state-of-the-art MARL algorithms in complex fully cooperative tasks from the MPE, LBF, and RWARE benchmarks.

Interactive Imitation Learning (IIL) allows agents to acquire desired behaviors through human interventions, but current methods impose high cognitive demands on human supervisors. We propose the Adaptive Intervention Mechanism (AIM), a novel robot-gated IIL algorithm that learns an adaptive criterion for requesting human demonstrations. AIM utilizes a proxy Q-function to mimic the human intervention rule and adjusts intervention requests based on the alignment between agent and human actions. By assigning high Q-values when the agent deviates from the expert and decreasing these values as the agent becomes proficient, the proxy Q-function enables the agent to assess the real-time alignment with the expert and request assistance when needed. Our expert-in-the-loop experiments reveal that AIM significantly reduces expert monitoring efforts in both continuous and discrete control tasks. Compared to the uncertainty-based baseline Thrifty-DAgger, our method achieves a 40% improvement in terms of human take-over cost and learning efficiency. Furthermore, AIM effectively identifies safety-critical states for expert assistance, thereby collecting higher-quality expert demonstrations and reducing overall expert data and environment interactions needed. Code and demo video are available at https://github.com/metadriverse/AIM.

Offline multi-agent reinforcement learning (MARL) struggles to estimate out-of-distribution states and actions due to the absence of real-time environmental feedback. While diffusion models show promise in addressing these challenges, their application primarily focuses on independently diffusing the historical trajectories of individual agents, neglecting crucial multi-agent coordination dynamics and reducing policy robustness in dynamic environments. In this paper, we propose MCGD, a novel Multi-agent Coordination framework based on Graph Diffusion models to improve the effectiveness and robustness of collaborative policies. Specifically, we begin by constructing a sparse coordination graph that includes continuous node attributes and discrete edge attributes to effectively identify the underlying dynamics of multi-agent interactions. Next, we derive transition probabilities between edge categories and present adaptive categorical diffusion to capture the structure diversity of multi-agent coordination. Leveraging this coordination structure, we define neighbor-dependent forward noise and develop anisotropic diffusion to enhance the action diversity of each agent. Extensive experiments across various multi-agent environments demonstrate that MCGD significantly outperforms existing state-of-the-art baselines in coordination performance and policy robustness in dynamic environments.

As autonomous systems become more ubiquitous in daily life, ensuring high performance with guaranteed safety is crucial. However, safety and performance could be competing objectives, which makes their co-optimization difficult. Learning-based methods, such as Constrained Reinforcement Learning (CRL), achieve strong performance but lack formal safety guarantees due to safety being enforced as soft constraints, limiting their use in safety-critical settings. Conversely, formal methods such as Hamilton-Jacobi (HJ) Reachability Analysis and Control Barrier Functions (CBFs) provide rigorous safety assurances but often neglect performance, resulting in overly conservative controllers. To bridge this gap, we formulate the co-optimization of safety and performance as a state-constrained optimal control problem, where performance objectives are encoded via a cost function and safety requirements are imposed as state constraints. We demonstrate that the resultant value function satisfies a Hamilton-Jacobi-Bellman (HJB) equation, which we approximate efficiently using a novel physics-informed machine learning framework. In addition, we introduce a conformal prediction-based verification strategy to quantify the learning errors, recovering a high-confidence safety value function, along with a probabilistic error bound on performance degradation. Through several case studies, we demonstrate the efficacy of the proposed framework in enabling scalable learning of safe and performant controllers for complex, high-dimensional autonomous systems.

The success of sequential decision-making approaches, such as *reinforcement learning* (RL), is closely tied to the availability of a reward feedback. However, designing a reward function that encodes the desired objective is a challenging task. In this work, we address a more realistic scenario: sequential decision making with preference feedback provided, for instance, by a human expert. We aim to build a theoretical basis linking *preferences*, (non-Markovian) *utilities*, and (Markovian) *rewards*, and we study the connections between them. First, we model preference feedback using a partial (pre)order over trajectories, enabling the presence of incomparabilities that are common when preferences are provided by humans but are surprisingly overlooked in existing works. Second, to provide a theoretical justification for a common practice, we investigate how a preference relation can be approximated by a multi-objective utility. We introduce a notion of preference-utility compatibility and analyze the computational complexity of this transformation, showing that constructing the minimum-dimensional utility is NP-hard. Third, we propose a novel concept of preference-based policy dominance that does not rely on utilities or rewards and discuss the computational complexity of assessing it. Fourth, we develop a computationally efficient algorithm to approximate a utility using (Markovian) rewards and quantify the error in terms of the suboptimality of the optimal policy induced by the approximating reward. This work aims to lay the foundation for a principled approach to sequential decision making from preference feedback, with promising potential applications in RL from human feedback.