ICLR 2026 论文列表

Offline reinforcement learning (RL) enables learning effective policies from fixed datasets without any environment interaction. Existing methods typically employ policy constraints to mitigate the distribution shift encountered during offline RL training. However, because the scale of the constraints varies across tasks and datasets of differing quality, existing methods must meticulously tune hyperparameters to match each dataset, which is time-consuming and often impractical. To bridge this gap, we propose Adaptive Scaling of Policy Constraints (ASPC), a second-order differentiable framework that automatically adjusts the scale of policy constraints during training. We theoretically analyze its performance improvement guarantee. In experiments on 39 datasets across four D4RL domains, ASPC using a single hyperparameter configuration outperforms other adaptive constraint methods and state-of-the-art offline RL algorithms that require per-dataset tuning, achieving an average 35\% improvement in normalized performance over the baseline. Moreover, ASPC consistently yields additional gains when integrated with a variety of existing offline RL algorithms, demonstrating its broad generality.

We propose FlowRL: matching the full reward distribution via flow balancing instead of solely maximizing rewards in large language model (LLM) reinforcement learning (RL). Recent advanced reasoning models adopt reward-maximizing methods (e.g., PPO and GRPO), which tend to over-optimize dominant reward signals while neglecting less frequent but valid reasoning paths, thus reducing diversity. In contrast, we transform scalar rewards into a normalized target distribution using a learnable partition function, and then minimize the reverse KL divergence between the policy and the target distribution. We implement this idea as a flow-balanced optimization method that promotes diverse exploration and generalizable reasoning trajectories. We conduct experiments on both math and code reasoning tasks: FlowRL achieves a significant average improvement of $10.0\%$ over GRPO and $5.1\%$ over PPO on math benchmarks, and performs consistently better on code reasoning tasks. These results highlight reward distribution-matching as a key step toward efficient exploration and diverse reasoning in LLM reinforcement learning.

The incorporation of memory into agents is essential for numerous tasks within the domain of Reinforcement Learning (RL). In particular, memory is paramount for tasks that require the use of past information, adaptation to novel environments, and improved sample efficiency. However, the term ``memory'' encompasses a wide range of concepts, which, coupled with the lack of a unified methodology for validating an agent's memory, leads to erroneous judgments about agents' memory capabilities and prevents objective comparison with other memory-enhanced agents. This paper aims to streamline the concept of memory in RL by providing practical precise definitions of agent memory types, such as long-term vs. short-term memory and declarative vs. procedural memory, inspired by cognitive science. Using these definitions, we categorize different classes of agent memory, propose a robust experimental methodology for evaluating the memory capabilities of RL agents, and standardize evaluations. Furthermore, we empirically demonstrate the importance of adhering to the proposed methodology when evaluating different types of agent memory by conducting experiments with different RL agents and what its violation leads to.

Exogenous MDPs (Exo-MDPs) capture sequential decision-making where uncertainty comes solely from exogenous inputs that evolve independently of the learner’s actions. This structure is especially common in operations research applications such as inventory control, energy storage, and resource allocation, where exogenous randomness (e.g., demand, arrivals, or prices) drives system behavior. Despite decades of empirical evidence that greedy, exploitation-only methods work remarkably well in these settings, theory has lagged behind: all existing regret guarantees for Exo-MDPs rely on explicit exploration or tabular assumptions. We show that exploration is unnecessary. We propose Pure Exploitation Learning ($\texttt{PEL}$) and prove the first general finite-sample regret bounds for exploitation-only algorithms in Exo-MDPs. In the tabular case, PEL achieves $\widetilde{O}(H^2|\Xi|\sqrt{K})$. For large, continuous endogenous state spaces, we introduce $\texttt{LSVI-PE}$, a simple linear-approximation method whose regret is polynomial in the feature dimension, exogenous state space, and horizon, independent of the endogenous state and action spaces. Our analysis introduces two new tools: counterfactual trajectories and Bellman-closed feature transport, which together allow greedy policies to have accurate value estimates without optimism. Experiments on synthetic and resource-management tasks show $\texttt{PEL}$ consistently outperforming baselines. Overall, our results overturn the conventional wisdom that exploration is required, demonstrating that in Exo-MDPs, pure exploitation is enough.

Sample efficiency is one of the main bottlenecks for optimal decision making via reinforcement learning. Pretraining a policy to maximize the entropy of the state visitation can substantially speedup reinforcement learning of downstream tasks. It is still an open question how to maximize the state entropy in POMDPs, where the true states of the environment, or their entropy, are not observed. In this work, we propose to maximize the entropy of a sufficient statistic of the history, which is called an information state. First, we show that a recursive latent model that predicts future observations is an information state in this setting. Then, we provide a practical algorithm, called LatEnt, to simultaneously learn the latent model and a latent-based policy maximizing the corresponding entropy objective from reward-free interactions with the POMDP. We empirically show that our approach induces higher state entropy than existing methods, which translates to better performance on downstream tasks. As a byproduct, we open-source PROBE, the first benchmark to test reward-free pretraining in POMDPs.

Reinforcement learning (RL) is a human-designed framework for solving sequential decision-making problems. In this work, we demonstrate that, surprisingly, RL emerges in LLMs at inference time – a phenomenon known as in-context RL (ICRL). To reveal this capability, we introduce a simple multi-round prompting framework, called ICRL prompting. The goal of ICRL prompting is to guide LLMs to perform reinforcement learning for self-improvement on a given task. After the LLM generates a response at the current round, we give numerical scalar feedback on the response, called the rewards. At the next round, we prompt the LLM again with the same task and a context consisting of all previous responses and rewards. We observe that the quality of the LLM's response increases as the context grows. In other words, the LLM is able to maximize the scalar reward signal at inference time, just like an RL algorithm. We evaluate ICRL prompting on Game of 24, creative writing, ScienceWorld, and Olympiad-level math competitions (AIME and HMMT), demonstrating significant improvements over baselines such as Self-Refine and Reflexion. Surprisingly, in some experiments, the reward signals are generated by the LLM itself, yet performance improvements are still observed from ICRL prompting, offering a new paradigm for test-time scaling.

Cross-domain reinforcement learning (CDRL) is meant to improve the data efficiency of RL by leveraging the data samples collected from a source domain to facilitate the learning in a similar target domain. Despite its potential, cross-domain transfer in RL is known to have two fundamental and intertwined challenges: (i) The source and target domains can have distinct state space or action space, and this makes direct transfer infeasible and thereby requires more sophisticated inter-domain mappings; (ii) The transferability of a source-domain model in RL is not easily identifiable a priori, and hence CDRL can be prone to negative effect during transfer. In this paper, we propose to jointly tackle these two challenges through the lens of \textit{cross-domain Bellman consistency} and \textit{hybrid critic}. Specifically, we first introduce the notion of cross-domain Bellman consistency as a way to measure transferability of a source-domain model. Then, we propose $Q$Avatar, which combines the Q functions from both the source and target domains with an adaptive hyperparameter-free weight function. Through this design, we characterize the convergence behavior of $Q$Avatar and show that $Q$Avatar achieves reliable transfer in the sense that it effectively leverages a source-domain Q function for knowledge transfer to the target domain. Through experiments, we demonstrate that $Q$Avatar achieves favorable transferability across various RL benchmark tasks, including locomotion and robot arm manipulation.

Multi-agent reinforcement learning (MARL) confronts a fundamental tension between stability and expressiveness. Stability requires avoiding divergence under non-stationary updates, while expressiveness demands capturing multimodal strategies for heterogeneous coordination. Existing methods sacrifice one for the other: value-decomposition and trust-region approaches ensure stability but assume restrictive unimodal policies, while expressive generative models lack optimization guarantees. To address this challenge, we introduce **H**eterogeneous **A**gent **Q**-weighted Policy **O**ptimization (**HAQO**), a framework unifying sequential advantage-aware updates, Q-weighted variational surrogates, and entropy regularization. Our analysis establishes monotone improvement guarantees under bounded critic bias, extending trust-region theory to diffusion-based policies with intractable log-likelihoods. HAQO achieves superior returns and reduced variance compared to policy-gradient baselines across diverse benchmarks. The ablation studies confirm sequential updates ensure stability, expressive policies enable multimodality, and entropy regularization prevents collapse. HAQO reconciles stability and expressiveness in MARL with theoretical rigor and practical effectiveness.

Controlling high-dimensional systems in biological and robotic applications is challenging due to expansive state–action spaces, where effective exploration is critical. Commonly used exploration strategies in reinforcement learning are largely undirected with sharp degradation as action dimensionality grows. Many existing methods resort to dimensionality reduction, which constrains policy expressiveness and forfeits system flexibility. We introduce Q-guided Flow Exploration (Qflex), a scalable reinforcement learning method that conducts exploration directly in the native high-dimensional action space. During training, Qflex traverses actions from a learnable source distribution along a probability flow induced by the learned value function, aligning exploration with task-relevant gradients rather than isotropic noise. Our proposed method substantially outperforms representative online reinforcement learning baselines across diverse high-dimensional continuous-control benchmarks. Qflex also successfully controls a full-body human musculoskeletal model to perform agile, complex movements, demonstrating superior scalability and sample efficiency in very high-dimensional settings. Our results indicate that value-guided flows offer a principled and practical route to exploration at scale.

The challenge of coalition formation games lies in efficiently navigating the exponentially large space of possible coalitions to identify the optimal partition. While existing approaches to solve coalition formation games either provide optimal solutions with limited scalability or approximate solutions without quality guarantees, we propose a novel scalable and sample-efficient approximation method based on deep reinforcement learning. Specifically, we model the coalition formation problem as a finite Markov decision process and use deep neural network to approximate optimal coalition structures within the full and abstracted coalition space. Moreover, our method is applicable to bi-level optimization problems in which coalition values are determined by the policies of individual agents at a lower decision-making level. This way, our approach facilitates dynamic, adaptive adjustments to coalition value assessments as they evolve over time. Empirical results demonstrate our algorithm's effectiveness in approximating optimal coalition structures in both normal-form and sequential mixed-motive games.

Transformers have become increasingly popular in offline reinforcement learning (RL) due to their ability to treat agent trajectories as sequences, reframing policy learning as a sequence modeling task. However, in partially observable environments (POMDPs), effective decision-making depends on retaining information about past events - something that standard transformers struggle with due to the quadratic complexity of self-attention, which limits their context length. One solution to this problem is to extend transformers with memory mechanisms. We propose the Recurrent Action Transformer with Memory (RATE), a novel transformer-based architecture for offline RL that incorporates a recurrent memory mechanism designed to regulate information retention. We evaluate RATE across a diverse set of environments: memory-intensive tasks (ViZDoom-Two-Colors, T-Maze, Memory Maze, Minigrid-Memory, and POPGym), as well as standard Atari and MuJoCo benchmarks. Our comprehensive experiments demonstrate that RATE significantly improves performance in memory-dependent settings while remaining competitive on standard tasks across a broad range of baselines. These findings underscore the pivotal role of integrated memory mechanisms in offline RL and establish RATE as a unified, high-capacity architecture for effective decision-making over extended horizons.

Dexterous manipulation is limited by both control and design, without consensus as to what makes manipulators best for performing dexterous tasks. This raises a fundamental challenge: how should we design and control robot manipulators that are optimized for dexterity? We present a co-design framework that learns task-specific hand morphology and complementary dexterous control policies. The framework supports 1) an expansive morphology search space including joint, finger, and palm generation, 2) scalable evaluation across the wide design space via morphology-conditioned cross-embodied control, and 3) real-world fabrication with accessible components. We evaluate the approach across multiple dexterous tasks, including in-hand rotation with simulation and real deployment. Our framework enables an end-to-end pipeline that can design, train, fabricate, and deploy a new robotic hand in under 24 hours. The full framework will be open-sourced and available on our website.

Unsupervised reinforcement learning (URL) pre-trains agents by exploring diverse states in reward-free environments, aiming to enable efficient adaptation to various downstream tasks. Without extrinsic rewards, prior methods rely on intrinsic objectives, but heterogeneous exploration data demand strong modeling capacity for both intrinsic reward design and policy learning. We introduce the **Ex**ploratory **D**iffusion **M**odel (**ExDM**), which leverages the expressive power of diffusion models to fit diverse replay-buffer distributions, thus providing accurate density estimates and a score-based intrinsic reward that drives exploration into under-visited regions. This mechanism substantially broadens state coverage and yields robust pre-trained policies. Beyond exploration, ExDM offers theoretical guarantees and practical algorithms for fine-tuning diffusion policies under limited interactions, overcoming instability and computational overhead from multi-step sampling. Extensive experiments on Maze2d and URLB show that ExDM achieves superior exploration and faster downstream adaptation, establishing new state-of-the-art results, particularly in environments with complex structure or cross-embodiment settings.

While reinforcement learning methods such as Group Relative Preference Optimization (GRPO) have significantly enhanced Large Language Models, adapting them to diffusion models remains challenging. In particular, GRPO demands a stochastic policy, yet the most cost‑effective diffusion samplers are based on deterministic ODEs. Recent work addresses this issue by using inefficient SDE-based samplers to induce stochasticity, but this reliance on model-agnostic Gaussian noise leads to slow convergence. To resolve this conflict, we propose Direct Group Preference Optimization (DGPO), a new online RL algorithm that dispenses with the policy-gradient framework entirely. DGPO learns directly from group-level preferences, which utilize relative information of samples within groups. This design eliminates the need for inefficient stochastic policies, unlocking the use of efficient deterministic ODE samplers and faster training. Extensive results show that DGPO trains around 20 times faster than existing state-of-the-art methods and achieves superior performance on both in-domain and out-of-domain reward metrics.

Behavioral Foundation Models (BFMs) have been recently successful in producing agents with the capabilities to adapt to any unknown reward or task. In reality, these methods are only able to produce near-optimal policies for the reward functions that are in the span of some pre-existing _state features_. Naturally, their efficiency relies heavily on the choice of state features that they use. As a result, these BFMs have used a wide variety of complex objectives, often sensitive to environment coverage, to train task spanning features with different inductive properties. With this work, our aim is to examine the question: are these complex representation learning objectives necessary for zero-shot RL? Specifically, we revisit the objective of self-supervised next-state prediction in latent space for state feature learning, but observe that such an objective alone is prone to increasing state-feature similarity, and subsequently reducing span of reward functions that we can represent optimal policies for. We propose an approach, RLDP, that adds a simple regularization to maintain feature diversity and can match or surpass state-of-the-art complex representation learning methods for zero-shot RL. Furthermore, we demonstrate the prior approaches diverge in low-coverage scenarios where RLDP still succeeds.

The Successor Measure (SM), a powerful method in reinforcement learning (RL), describes discounted future state distributions under a policy, and it has recently been studied using generative modeling techniques. Although SM is a powerful predictive object, it lacks compact representations tailored for online RL. To address this, we introduce Successor Flow Features (SF2), a representation learning framework that bridges SM estimation with policy optimization. SF2 leverages flow-matching generative models to approximate successor measures, while enforcing a structured linear decomposition into a time-invariant embedding and a time-dependent projection. This yields compact, policy-aware state-action features that integrate readily into standard off-policy algorithms like TD3 and SAC. Experiments on DeepMind Control Suite tasks show that SF2 improves sample efficiency and training stability compared to strong successor feature baselines. We attribute these gains to the compact representation induced by flow matching, which reduces compounding errors in long-horizon predictions.

Large-scale pre-training has fundamentally changed how machine learning research is done today: large foundation models are trained once, and then can be used by anyone in the community (including those without data or compute resources to train a model from scratch) to adapt and fine-tune to specific tasks. Applying this same framework to reinforcement learning (RL) is appealing because it offers compelling avenues for addressing core challenges in RL, including sample efficiency and robustness. However, there remains a fundamental challenge to pre-train large models in the context of RL: actions have long-term dependencies, so training a foundation model that reasons across *time* is important. Recent advances in generative AI have provided new tools for modeling highly complex distributions. In this paper, we build a probabilistic model to predict which states an agent will visit in the temporally distant future (i.e., an occupancy measure) using flow matching. As large datasets are often constructed by many distinct users performing distinct tasks, we include in our model a latent variable capturing the user intention. This intention increases the expressivity of our model, and enables adaptation with generalized policy improvement. We call our proposed method **intention-conditioned flow occupancy models (InFOM)**. Comparing with alternative methods for pre-training, our experiments on $36$ state-based and $4$ image-based benchmark tasks demonstrate that the proposed method achieves $1.8 \times$ median improvement in returns and increases success rates by $36\\%$.

Achieving optimality and adversarial robustness in deep reinforcement learning has long been regarded as conflicting goals. Nonetheless, recent theoretical insights presented in CAR suggest a potential alignment, raising the important question of how to realize this in practice. This paper first identifies a key gap between theory and practice by comparing standard policy optimization (SPO) and adversarially robust policy optimization (ARPO). Although they share theoretical consistency, *a fundamental tension between robustness and optimality arises in practical policy gradient methods*. SPO tends toward convergence to vulnerable first-order stationary policies (FOSPs) with strong natural performance, whereas ARPO typically favors more robust FOSPs at the expense of reduced returns. Furthermore, we attribute this tradeoff to the *reshaping effect of the strongest adversaries* in ARPO, which significantly complicates the global landscape by inducing *deceptive sticky FOSPs*. This improves robustness but makes navigation more challenging. To alleviate this, we develop the *BARPO*, a bilevel framework unifying SPO and ARPO by modulating adversary strength, thereby facilitating navigability while preserving global optima. Extensive empirical results demonstrate that BARPO consistently outperforms vanilla ARPO, providing a practical approach to reconcile theoretical and empirical performance.

Deep neural networks that incorporate classic reinforcement learning methods, such as value iteration, into their structure significantly outperform randomly structured networks in learning and generalization. These networks, however, are mostly limited to environments with no or very low uncertainty and do not extend well to partially observable environments. In this paper, we propose a new planning module architecture, the VI$^2$N (Value Iteration with Value of Information Network), that learns to act in novel environments with high perceptual ambiguity. This architecture over-emphasizes reducing uncertainty before exploiting the reward. VI$^2$N can also utilize factorization in environments with mixed observability to decrease the computational complexity of calculating the policy and to facilitate learning. Tested on a range of grid-based navigation tasks, each containing various types of environments with different degrees of observability, our network outperforms other deep architectures. Moreover, VI$^2$N generates interpretable cognitive maps highlighting both rewarding and informative locations. These maps highlight the key states the agent must visit to achieve its goal.