ICML 2025 论文列表

We present an approach to model-based RL that achieves a new state of the art performance on the challenging Craftax-classic benchmark, an open-world 2D survival game that requires agents to exhibit a wide range of general abilities---such as strong generalization, deep exploration, and long-term reasoning. With a series of careful design choices aimed at improving sample efficiency, our MBRL algorithm achieves a reward of 69.66% after only 1M environment steps, significantly outperforming DreamerV3, which achieves $53.2\%$, and, for the first time, exceeds human performance of 65.0%. Our method starts by constructing a SOTA model-free baseline, using a novel policy architecture that combines CNNs and RNNs. We then add three improvements to the standard MBRL setup: (a) "Dyna with warmup", which trains the policy on real and imaginary data, (b) "nearest neighbor tokenizer" on image patches, which improves the scheme to create the transformer world model (TWM) inputs, and (c) "block teacher forcing", which allows the TWM to reason jointly about the future tokens of the next timestep.

We present a Transformer-based framework for Constraint Satisfaction Problems (CSPs). CSPs find use in many applications and thus accelerating their solution with machine learning is of wide interest. Most existing approaches rely on supervised learning from feasible solutions or reinforcement learning, paradigms that require either feasible solutions to these NP-Complete CSPs or large training budgets and a complex expert-designed reward signal. To address these challenges, we propose ConsFormer, a self-supervised framework that leverages a Transformer as a solution refiner. ConsFormer constructs a solution to a CSP iteratively in a process that mimics local search. Instead of using feasible solutions as labeled data, we devise differentiable approximations to the discrete constraints of a CSP to guide model training. Our model is trained to improve random assignments for a single step but is deployed iteratively at test time, circumventing the bottlenecks of supervised and reinforcement learning. Experiments on Sudoku, Graph Coloring, Nurse Rostering, and MAXCUT demonstrate that our method can tackle out-of-distribution CSPs simply through additional iterations.

We explore the potential of \emph{AI-enhanced combinatorial optimization theory}, taking online bipartite matching (OBM) as a case study. In the theoretical study of OBM, the \emph{hardness} corresponds to a performance \emph{upper bound} of a specific online algorithm or any possible online algorithms. Typically, these upper bounds derive from challenging instances meticulously designed by theoretical computer scientists. Zhang et al. (ICML 2024) recently provide an example demonstrating how reinforcement learning techniques enhance the hardness result of a specific OBM model. Their attempt is inspiring but preliminary. It is unclear whether their methods can be applied to other OBM problems with similar breakthroughs. This paper takes a further step by introducing DiMa, a unified and novel framework that aims at understanding the hardness of OBM problems based on denoising diffusion probabilistic models (DDPMs). DiMa models the process of generating hard instances as denoising steps, and optimizes them by a novel reinforcement learning algorithm, named \emph{shortcut policy gradient} (SPG). We first examine DiMa on the classic OBM problem by reproducing its known hardest input instance in literature. Further, we apply DiMa to two well-known variants of OBM, for which the exact hardness remains an open problem, and we successfully improve their theoretical state-of-the-art upper bounds.

The general-utility Markov decision processes (GUMDPs) framework generalizes the MDPs framework by considering objective functions that depend on the frequency of visitation of state-action pairs induced by a given policy. In this work, we contribute with the first analysis on the impact of the number of trials, i.e., the number of randomly sampled trajectories, in infinite-horizon GUMDPs. We show that, as opposed to standard MDPs, the number of trials plays a key-role in infinite-horizon GUMDPs and the expected performance of a given policy depends, in general, on the number of trials. We consider both discounted and average GUMDPs, where the objective function depends, respectively, on discounted and average frequencies of visitation of state-action pairs. First, we study policy evaluation under discounted GUMDPs, proving lower and upper bounds on the mismatch between the finite and infinite trials formulations for GUMDPs. Second, we address average GUMDPs, studying how different classes of GUMDPs impact the mismatch between the finite and infinite trials formulations. Third, we provide a set of empirical results to support our claims, highlighting how the number of trajectories and the structure of the underlying GUMDP influence policy evaluation.

Off-policy learning and evaluation leverage logged bandit feedback datasets, which contain context, action, propensity score, and feedback for each data point. These scenarios face significant challenges due to high variance and poor performance with low-quality propensity scores and heavy-tailed reward distributions. We address these issues by introducing a novel estimator based on the log-sum-exponential (LSE) operator, which outperforms traditional inverse propensity score estimators. Our LSE estimator demonstrates variance reduction and robustness under heavy-tailed conditions. For off-policy evaluation, we derive upper bounds on the estimator's bias and variance. In the off-policy learning scenario, we establish bounds on the regret—the performance gap between our LSE estimator and the optimal policy—assuming bounded $(1+\epsilon)$-th moment of weighted reward. Notably, we achieve a convergence rate of $O(n^{-\epsilon/(1+\epsilon)})$ for the regret bounds, where $\epsilon\in[0,1]$ and $n$ is the size of logged bandit feedback dataset. Theoretical analysis is complemented by comprehensive empirical evaluations in both off-policy learning and evaluation scenarios, confirming the practical advantages of our approach. The code for our estimator is available at the following link: https://github.com/armin-behnamnia/lse-offpolicy-learning .

Reward shaping has been demonstrated to be an effective technique for accelerating the learning process of reinforcement learning (RL) agents. While successful in empirical applications, the design of a good shaping function is less well understood in principle and thus often relies on domain expertise and manual design. To overcome this limitation, we propose a novel automated approach for designing reward functions from offline data, possibly contaminated with the unobserved confounding bias.We propose to use causal state value upper bounds calculated from offline datasets as a conservative optimistic estimation of the optimal state value, which is then used as state potentials in Potential-Based Reward Shaping (PBRS). When applying our shaping function to a model-free learner based on UCB principles, we show that it enjoys a better gap-dependent regret bound than the learner without shaping. To the best of our knowledge, this is the first gap-dependent regret bound for PBRS in model-free learning with online exploration. Simulations support the theoretical findings.

Online fine-tuning vision-language model (VLM) agents with reinforcement learning (RL) has shown promise for equipping agents with multi-step, goal-oriented capabilities in dynamic environments. However, their open-ended textual action space and non-end-to-end nature of action generation present significant challenges to effective online exploration in RL, e.g., explosion of the exploration space. We propose a novel online fine-tuning method, Counterfactual Soft Reinforcement Learning (CoSo), better suited to the textual output space of VLM agents. Compared to prior methods that assign uniform uncertainty to all tokens, CoSo leverages counterfactual reasoning to dynamically assess the causal influence of individual tokens on post-processed actions. By prioritizing the exploration of action-critical tokens while reducing the impact of semantically redundant or low-impact tokens, CoSo enables a more targeted and efficient online rollout process. We provide theoretical analysis proving CoSo's convergence and policy improvement guarantees, and extensive empirical evaluations supporting CoSo's effectiveness. Our results across a diverse set of agent tasks, including Android device control, card gaming, and embodied AI, highlight its remarkable ability to enhance exploration efficiency and deliver consistent performance gains. The code is available at https://github.com/langfengQ/CoSo.

To learn from data collected in diverse dynamics, Imitation from Observation (IfO) methods leverage expert state trajectories based on the premise that recovering expert state distributions in other dynamics facilitates policy learning in the current one. However, Imitation Learning inherently imposes a performance upper bound of learned policies. Additionally, as the environment dynamics change, certain expert states may become inaccessible, rendering their distributions less valuable for imitation. To address this, we propose a novel framework that integrates reward maximization with IfO, employing F-distance regularized policy optimization. This framework enforces constraints on globally accessible states—those with nonzero visitation frequency across all considered dynamics—mitigating the challenge posed by inaccessible states. By instantiating F-distance in different ways, we derive two theoretical analysis and develop a practical algorithm called Accessible State Oriented Policy Regularization (ASOR). ASOR serves as a general-purpose module that can be incorporated into various RL approaches, including offline RL and off-policy RL. Extensive experiments across multiple benchmarks demonstrate ASOR’s effectiveness in enhancing state-of-the-art cross-domain policy transfer algorithms, significantly improving their performance.

A tenet of reinforcement learning is that the agent always observes rewards. However, this is not true in many realistic settings, e.g., a human observer may not always be available to provide rewards, sensors may be limited or malfunctioning, or rewards may be inaccessible during deployment. Monitored Markov decision processes (Mon-MDPs) have recently been proposed to model such settings. However, existing Mon-MDP algorithms have several limitations: they do not fully exploit the problem structure, cannot leverage a known monitor, lack worst-case guarantees for "unsolvable" Mon-MDPs without specific initialization, and offer only asymptotic convergence proofs. This paper makes three contributions. First, we introduce a model-based algorithm for Mon-MDPs that addresses these shortcomings. The algorithm employs two instances of model-based interval estimation: one to ensure that observable rewards are reliably captured, and another to learn the minimax-optimal policy. Second, we empirically demonstrate the advantages. We show faster convergence than prior algorithms in more than four dozen benchmarks, and even more dramatic improvements when the monitoring process is known. Third, we present the first finite-sample bound on performance. We show convergence to a minimax-optimal policy even when some rewards are never observable.

Building neural reward models from human preferences is a pivotal component in reinforcement learning from human feedback (RLHF) and large language model alignment research. Given the scarcity and high cost of human annotation, how to select the most informative pairs to annotate is an essential yet challenging open problem. In this work, we highlight the insight that an ideal comparison dataset for reward modeling should balance *exploration of the representation space* and make *informative comparisons* between pairs with moderate reward differences. Technically, challenges arise in quantifying the two objectives and efficiently prioritizing the comparisons to be annotated. To address this, we propose the Fisher information-based selection strategies, adapt theories from the *classical experimental design* literature, and apply them to the final linear layer of the deep neural network-based reward modeling tasks. Empirically, our method demonstrates remarkable performance, high computational efficiency, and stability compared to other selection methods from deep learning and classical statistical literature across multiple open-source LLMs and datasets. Further ablation studies reveal that incorporating cross-prompt comparisons in active reward modeling significantly enhances labeling efficiency, shedding light on the potential for improved annotation strategies in RLHF. Code and embeddings to reproduce all results of this paper are available at https://github.com/YunyiShen/ARM-FI/.

Generative Flow Networks (GFNs) were initially introduced on directed non-acyclic graphs to sample from an unnormalized distribution density. Recent works have extended the theoretical framework for generative methods allowing more flexibility and enhancing application range. However, many challenges remain in training GFNs in continuous settings and for imitation learning (IL), including intractability of flow-matching loss, limited tests of non-acyclic training, and the need for a separate reward model in imitation learning. The present work proposes a family of generative flows called Ergodic Generative Flows (EGFs) which are used to address the aforementioned issues. First, we leverage ergodicity to build simple generative flows with finitely many globally defined transformations (diffeomorphisms) with universality guarantees and tractable flow-matching loss (FM loss). Second, we introduce a new loss involving cross-entropy coupled to weak flow-matching control, coined KL-weakFM loss. It is designed for IL training without a separate reward model. We evaluate IL-EGFs on toy 2D tasks and real-world datasets from NASA on the sphere, using the KL-weakFM loss. Additionally, we conduct toy 2D reinforcement learning experiments with a target reward, using the FM loss.

Recently, incorporating knowledge from pretrained language models (PLMs) into decision transformers (DTs) has generated significant attention in offline reinforcement learning (RL). These PLMs perform well in RL tasks, raising an intriguing question: what kind of knowledge from PLMs has been transferred to RL to achieve such good results? This work first dives into this problem by analyzing each head quantitatively and points out Markov head, a crucial component that exists in the attention heads of PLMs. It leads to extreme attention on the last-input token and performs well only in short-term environments. Furthermore, we prove that this extreme attention cannot be changed by re-training embedding layer or fine-tuning. Inspired by our analysis, we propose a general method GPT-DTMA, which equips a pretrained DT with Mixture of Attention (MoA), to enable adaptive learning and accommodate diverse attention requirements during fine-tuning. Extensive experiments demonstrate the effectiveness of GPT-DTMA: it achieves superior performance in short-term environments compared to baselines, significantly reduces the performance gap of PLMs in long-term scenarios, and the experimental results also validate our theorems.

The Robust Regularized Markov Decision Process (RRMDP) is proposed to learn policies robust to dynamics shifts by adding regularization to the transition dynamics in the value function. Existing methods mostly use unstructured regularization, potentially leading to conservative policies under unrealistic transitions. To address this limitation, we propose a novel framework, the $d$-rectangular linear RRMDP ($d$-RRMDP), which introduces latent structures into both transition kernels and regularization. We focus on offline reinforcement learning, where an agent learns policies from a precollected dataset in the nominal environment. We develop the Robust Regularized Pessimistic Value Iteration (R2PVI) algorithm that employs linear function approximation for robust policy learning in $d$-RRMDPs with $f$-divergence based regularization terms on transition kernels. We provide instance-dependent upper bounds on the suboptimality gap of R2PVI policies, demonstrating that these bounds are influenced by how well the dataset covers state-action spaces visited by the optimal robust policy under robustly admissible transitions. We establish information-theoretic lower bounds to verify that our algorithm is near-optimal. Finally, numerical experiments validate that R2PVI learns robust policies and exhibits superior computational efficiency compared to baseline methods.

Achieving seamless coordination between AI agents and humans is crucial for real-world applications, yet it remains a significant open challenge. Hanabi is a cooperative card game featuring imperfect information, constrained communication, theory of mind requirements, and coordinated action -- making it an ideal testbed for human-AI coordination. However, its use for human-AI interaction has been limited by the challenges of human evaluation. In this work, we introduce the Ad-Hoc Human-AI Coordination Challenge (AH2AC2) to overcome the constraints of costly and difficult-to-reproduce human evaluations. We develop \textit{human proxy agents} on a large-scale human dataset that serve as robust, cheap, and reproducible human-like evaluation partners in AH2AC2. To encourage the development of data-efficient methods, we open-source a dataset of 3,079 games, deliberately limiting the amount of available human gameplay data. We present baseline results for both two- and three- player Hanabi scenarios. To ensure fair evaluation, we host the proxy agents through a controlled evaluation system rather than releasing them publicly. The code is available at \href{https://github.com/FLAIROx/ah2ac2}{https://github.com/FLAIROx/ah2ac2}.

We study online finite-horizon Markov Decision Processes with adversarially changing loss and aggregate bandit feedback (a.k.a full-bandit). Under this type of feedback, the agent observes only the total loss incurred over the entire trajectory, rather than the individual losses at each intermediate step within the trajectory. We introduce the first Policy Optimization algorithms for this setting. In the known-dynamics case, we achieve the first *optimal* regret bound of $\tilde \Theta(H^2\sqrt{SAK})$, where $K$ is the number of episodes, $H$ is the episode horizon, $S$ is the number of states, and $A$ is the number of actions. In the unknown dynamics case we establish regret bound of $\tilde O(H^3 S \sqrt{AK})$, significantly improving the best known result by a factor of $H^2 S^5 A^2$.

This paper introduces von Mises-Fisher exploration (vMF-exp), a scalable method for exploring large action sets in reinforcement learning problems where hyperspherical embedding vectors represent these actions. vMF-exp involves initially sampling a state embedding representation using a von Mises-Fisher distribution, then exploring this representation's nearest neighbors, which scales to virtually unlimited numbers of candidate actions. We show that, under theoretical assumptions, vMF-exp asymptotically maintains the same probability of exploring each action as Boltzmann Exploration (B-exp), a popular alternative that, nonetheless, suffers from scalability issues as it requires computing softmax values for each action. Consequently, vMF-exp serves as a scalable alternative to B-exp for exploring large action sets with hyperspherical embeddings. Experiments on simulated data, real-world public data, and the successful large-scale deployment of vMF-exp on the recommender system of a global music streaming service empirically validate the key properties of the proposed method.

Inferring reward functions from demonstrations is a key challenge in reinforcement learning (RL), particularly in multi-agent RL (MARL). The large joint state-action spaces and intricate inter-agent interactions in MARL make inferring the joint reward function especially challenging. While prior studies in single-agent settings have explored ways to recover reward functions and expert policies from human preference feedback, such studies in MARL remain limited. Existing methods typically combine two separate stages, supervised reward learning, and standard MARL algorithms, leading to unstable training processes. In this work, we exploit the inherent connection between reward functions and Q functions in cooperative MARL to introduce a novel end-to-end preference-based learning framework. Our framework is supported by a carefully designed multi-agent value decomposition strategy that enhances training efficiency. Extensive experiments on two state-of-the-art benchmarks, SMAC and MAMuJoCo, using preference data generated by both rule-based and large language model approaches demonstrate that our algorithm consistently outperforms existing methods across various tasks.

Reinforcement learning with offline data suffers from Q-value extrapolation errors. To address this issue, we first demonstrate that linear extrapolation of the Q-function beyond the data range is particularly problematic. To mitigate this, we propose guiding the gradual decrease of Q-values outside the data range, which is achieved through reward scaling with layer normalization (RS-LN) and a penalization mechanism for infeasible actions (PA). By combining RS-LN and PA, we develop a new algorithm called PARS. We evaluate PARS across a range of tasks, demonstrating superior performance compared to state-of-the-art algorithms in both offline training and online fine-tuning on the D4RL benchmark, with notable success in the challenging AntMaze Ultra task.

The Maximum Entropy Reinforcement Learning (MaxEnt RL) framework is a leading approach for achieving efficient learning and robust performance across many RL tasks. However, MaxEnt methods have also been shown to struggle with performance-critical control problems in practice, where non-MaxEnt algorithms can successfully learn. In this work, we analyze how the trade-off between robustness and optimality affects the performance of MaxEnt algorithms in complex control tasks: while entropy maximization enhances exploration and robustness, it can also mislead policy optimization, leading to failure in tasks that require precise, low-entropy policies. Through experiments on a variety of control problems, we concretely demonstrate this misleading effect. Our analysis leads to better understanding of how to balance reward design and entropy maximization in challenging control problems.