ICLR 2026 Papers

Offline Multi-Agent Reinforcement Learning (MARL) enables policy learning from static datasets in multi-agent systems, eliminating the need for risky or costly environment interactions during training. A central challenge in offline MARL lies in achieving effective collaboration among heterogeneous agents under the constraints of fixed datasets, where \textbf{conservatism} is introduced to restrict behaviors to data-supported distributions. Agents with distinct roles and capabilities require individualized conservatism - yet must maintain cohesive team performance. However, existing approaches often apply uniform conservatism across all agents, leading to over-constraining critical agents and under-constraining others, which hampers effective collaboration. To address this issue, a novel framework, \textbf{OMCDA}, is proposed, where the degree of conservatism is dynamically adjusted for individual agents based on their impact on overall system performance. The framework is characterized by two key innovations: (1) A decomposed Q-function architecture is introduced to disentangle return computation from policy deviation assessment, allowing precise evaluations of each agent's contribution; and (2) An adaptive conservatism mechanism is developed to scale constraint strength according to both behavior policy divergence and the estimated importance of agents to the system. Experiments on MuJoCo and SMAC show OMCDA outperforms existing offline MARL methods, effectively balancing the flexibility and conservatism across agents while ensuring fair credit assignment and better collaboration.

Reinforcement learning (RL) agents often face high sample complexity in sparse or delayed reward settings, due to limited prior knowledge. Conversely, large language models (LLMs) can provide subgoal structures, plausible trajectories, and abstract priors that support early learning. Yet heavy reliance on LLMs introduces scalability issues and risks dependence on unreliable signals, motivating ongoing efforts to integrate LLM guidance without compromising RL’s autonomy. We propose MIRA (Memory-Integrated Reinforcement Learning Agent), which incorporates a structured, evolving memory graph to guide early learning. This graph stores decision-relevant information, such as trajectory segments and subgoal decompositions, and is co-constructed from the agent’s high-return experiences and LLM outputs, amortizing LLM queries into a persistent memory instead of relying on continuous real-time supervision. From this structure, we derive a utility signal that softly adjusts advantage estimation to refine policy updates without altering the underlying reward function. As training progresses, the agent’s policy surpasses the initial LLM-derived priors, and the utility term decays, leaving long-term convergence guarantees intact. We show theoretically that this utility-based shaping improves early-stage learning in sparse-reward settings. Empirically, MIRA outperforms RL baselines and reaches returns comparable to methods that rely on frequent LLM supervision, while requiring substantially fewer online LLM queries.

Single-domain offline reinforcement learning (RL) often suffers from limited data coverage, while cross-domain offline RL handles this issue by leveraging additional data from other domains with dynamics shifts. However, existing studies primarily focus on train-time robustness (handling dynamics shifts from training data), neglecting the test-time robustness against dynamics perturbations when deployed in practical scenarios. In this paper, we investigate dual (both train-time and test-time) robustness against dynamics shifts in cross-domain offline RL. We first empirically show that the policy trained with cross-domain offline RL exhibits fragility under dynamics perturbations during evaluation, particularly when target domain data is limited. To address this, we introduce a novel robust cross-domain Bellman (RCB) operator, which enhances test-time robustness against dynamics perturbations while staying conservative to the out-of-distribution dynamics transitions, thus guaranteeing the train-time robustness. To further counteract potential value overestimation or underestimation caused by the RCB operator, we introduce two techniques, the dynamic value penalty and the Huber loss, into our framework, resulting in the practical Dual-RObust Cross-domain Offline RL (DROCO) algorithm. Extensive empirical results across various dynamics shift scenarios show that DROCO outperforms strong baselines and exhibits enhanced robustness to dynamics perturbations.

Deploying reinforcement learning (RL) in robotics, industry, and health care is blocked by two obstacles: the difficulty of specifying accurate rewards and the risk of unsafe, data-hungry exploration. We address this by proposing a two-stage framework that first learns a safe initial policy from a reward-free dataset of expert demonstrations, then fine-tunes it online using preference-based human feedback. We provide the first principled analysis of this offline-to-online approach and introduce BRIDGE, a unified algorithm that integrates both signals via an uncertainty-weighted objective. We derive regret bounds that shrink with the number of offline demonstrations, explicitly connecting the quantity of offline data to online sample efficiency. We validate BRIDGE in discrete and continuous control MuJoCo environments, showing it achieves lower regret than both standalone behavioral cloning and online preference-based RL. Our work establishes a theoretical foundation for designing more sample-efficient interactive agents.

Leveraging offline data is a promising way to improve the sample efficiency of online reinforcement learning (RL). This paper expands the pool of usable data for offline-to-online RL by leveraging abundant non-curated data that is reward-free, of mixed quality, and collected across multiple embodiments. Although learning a world model appears promising for utilizing such data, we find that naive fine-tuning fails to accelerate RL training on many tasks. Through careful investigation, we attribute this failure to the distributional shift between offline and online data during fine-tuning. To address this issue and effectively use the offline data, we propose two techniques: i) experience rehearsal and ii) execution guidance. With these modifications, the non-curated offline data substantially improves RL’s sample efficiency. Under limited sample budgets, our method achieves a 102.8% relative improvement in aggregate score over learning-from-scratch baselines across 72 visuomotor tasks spanning 6 embodiments. On challenging tasks such as locomotion and robotic manipulation, it outperforms prior methods that utilize offline data by a decent margin.

The recent paradigm shift towards training large language models (LLMs) using DeepSeek-R1-Zero-style reinforcement learning (RL) on verifiable rewards has led to impressive advancements in code and mathematical reasoning. However, this methodology is limited to tasks where rule-based answer verification is possible and does not naturally extend to real-world domains such as chemistry, healthcare, engineering, law, biology, business, and economics. Current practical workarounds use an additional LLM as a model-based verifier; however, this introduces issues such as reliance on a strong verifier LLM, susceptibility to reward hacking, and the practical burden of maintaining the verifier model in memory during training. To address this and extend DeepSeek-R1-Zero-style training to general reasoning domains, we propose a verifier-free method (**VeriFree**) that bypasses answer verification and instead directly maximizes the probability of generating the reference answer, derived in a principled way from the RL objective. We compare VeriFree with verifier-based methods and demonstrate that, in addition to its significant practical benefits and reduced compute requirements, VeriFree matches and even surpasses verifier-based methods on extensive evaluations across MMLU-Pro, GPQA, SuperGPQA, and math-related benchmarks.

The expansion of generative AI services highlights the growing need for adaptive mechanisms to select an appropriate generative model for responding to a user's prompts. Recent works have proposed offline and online learning formulations to identify the optimal generative model for an input prompt, based solely on maximizing prompt-based fidelity evaluation scores, e.g., CLIP-Score in text-to-image generation. However, such fidelity-based selection methods overlook the diversity of generated outputs, and hence, they can fail to address potential diversity shortcomings in the generated responses. In this paper, we introduce the *Diversity-Aware Kernelized Upper Confidence Bound (DAK-UCB)* method as a contextual bandit algorithm for the online selection of generative models with diversity considerations. The proposed DAK-UCB method incorporates both fidelity and diversity-related metrics into the selection process. We design this framework based on prompt-aware diversity score functions that decompose to a two-sample-based expectation over prompt-output pairs in the previous generation rounds. Specifically, we illustrate the application of our framework using joint kernel distance and kernel entropy measures. Our experimental results demonstrate the effectiveness of DAK-UCB in promoting diversity-aware model selection while maintaining fidelity in the generations for a sequence of prompts.

Recent GRPO-based approaches built on flow matching models have shown remarkable improvements in human preference alignment for text-to-image generation. Nevertheless, they still suffer from the sparse reward problem: the terminal reward of the entire denoising trajectory is applied to all intermediate steps, resulting in a mismatch between the global feedback signals and the exact fine-grained contributions at intermediate denoising steps. To address this issue, we introduce \textbf{DenseGRPO}, a novel framework that aligns human preference with dense rewards, which evaluates the fine-grained contribution of each denoising step. Specifically, our approach includes two key components: (1) we propose to predict the step-wise reward gain as dense reward of each denoising step, which applies a reward model on the intermediate clean images via an ODE-based approach. This manner ensures an alignment between feedback signals and the contributions of individual steps, facilitating effective training; and (2) based on the estimated dense rewards, a mismatch drawback between the uniform exploration setting and the time-varying noise intensity in existing GRPO-based methods is revealed, leading to an inappropriate exploration space. Thus, we propose a reward-aware scheme to calibrate the exploration space by adaptively adjusting a timestep-specific stochasticity injection in the SDE sampler, ensuring a suitable exploration space at all timesteps. Extensive experiments on multiple standard benchmarks demonstrate the effectiveness of the proposed DenseGRPO and highlight the critical role of the valid dense rewards in flow matching model alignment.

Group Relative Policy Optimization (GRPO) has significantly enhanced the reasoning capability of large language models by optimizing the arithmetic mean of token-level rewards. Unfortunately, GRPO is observed to suffer from unstable policy updates when facing tokens with outlier importance-weighted rewards, which manifest as extreme importance sampling ratios during training. In this study, we propose Geometric-Mean Policy Optimization (GMPO), with the aim to improve the stability of GRPO through suppressing token reward outliers. GMPO is plug-and-play—simply replacing GRPO's arithmetic mean with the geometric mean of token-level rewards, as the latter is inherently less sensitive to outliers. GMPO is theoretically plausible—analysis reveals that both GMPO and GRPO are weighted forms of the policy gradient while the former enjoys more stable weights, which consequently benefits policy optimization and performance. Experiments on multiple mathematical reasoning benchmarks show that \Ours-7B improves the average Pass@1 of GRPO by up to 4.1%, outperforming many state-of-the-art approaches. Code is available at https://github.com/callsys/GMPO and verl.

Model-based reinforcement learning promises strong sample efficiency but often underperforms in practice due to compounding model error, unimodal world models that average over multi-modal dynamics, and overconfident predictions that bias learning. We introduce WIMLE, a model-based method that extends Implicit Maximum Likelihood Estimation (IMLE) to the model-based RL framework to learn stochastic, multi-modal world models without iterative sampling and to estimate predictive uncertainty via ensembles and latent sampling. During training, WIMLE weights each synthetic transition by its predicted confidence, preserving useful model rollouts while attenuating bias from uncertain predictions and enabling stable learning. Across 40 continuous-control tasks spanning DeepMind Control, MyoSuite, and HumanoidBench, WIMLE achieves superior sample efficiency and competitive or better asymptotic performance than strong model-free and model-based baselines. Notably, on the challenging Humanoid-run task, WIMLE improves sample efficiency by over 50\% relative to the strongest competitor, and on HumanoidBench it solves 8 of 14 tasks (versus 4 for BRO and 5 for SimbaV2). These results highlight the value of IMLE-based multi-modality and uncertainty-aware weighting for stable model-based RL.

In this work, we take a first step toward elucidating the mechanisms behind emergent exploration in unsupervised reinforcement learning. We study Single-Goal Contrastive Reinforcement Learning (SGCRL) (Liu et al., 2025), a self-supervised algorithm capable of solving challenging long-horizon goal-reaching tasks without external rewards or curricula. We combine theoretical analysis of the algorithm’s objective function with controlled experiments to understand what drives its exploration. We show that SGCRL maximizes implicit rewards shaped by its learned representations. These representations automatically modify the reward landscape to promote exploration before reaching the goal and exploitation thereafter. Our experiments also demonstrate that these exploration dynamics arise from learning low-rank representations of the state space rather than from neural network function approximation. Our improved understanding enables us to adapt SGCRL to perform safety-aware exploration.

Reinforcement post training (RPT) has recently shown promise in improving the reasoning abilities of large language models (LLMs). However, it remains unclear how well these improvements generalize to new domains, as prior work evaluates RPT models on data from the same domains used for fine-tuning. To understand the generalizability of RPT, we conduct two studies. (1) Observational: we compare a wide range of open-weight RPT models against their corresponding base models across multiple domains, including both seen and unseen domains in their fine-tuning data. (2) Interventional: we fine-tune LLMs with RPT on single domains and evaluate their performance across multiple domains. Both studies converge on the same conclusion that, although RPT brings substantial gains on tasks similar to the fine-tuning data, the gains generalize inconsistently and can vanish on domains with different reasoning patterns.

As single-center computing approaches power constraints, decentralized training becomes essential. However, traditional Reinforcement Learning (RL) methods, crucial for enhancing large model post-training, cannot adapt to decentralized distributed training due to the tight coupling between parameter learning and rollout sampling. For this, we propose HeteroRL, a heterogeneous RL architecture that decouples these processes, enabling stable training across geographically distributed nodes connected via the Internet. The core component is Group Expectation Policy Optimization (GEPO), an asynchronous RL algorithm robust to latency caused by network delays or heterogeneity in computational resources. Our study reveals that high latency significantly increases KL divergence, leading to higher variance of importance weights and training instability. GEPO mitigates this issue by using group expectation weighting to exponentially reduce the variance of importance weights, with theoretical guarantees. Experiments show GEPO achieves superior stability—only a 3\% performance drop from online to 1800s latency—and reduces the best-to-last gap by 85\% versus GSPO ($\Delta$=1.8 vs. 12.0) while attaining the highest scores, highlighting its effectiveness in decentralized, resource-heterogeneous environments.

Machine learning engineering (MLE) has a clear objective: Given an MLE task and a verifier (e.g., performance on some held-out data), what is the most effective way to utilize compute to achieve the best performance for the given task? Existing language model (LM) agents rely on prompting frontier LMs and accumulating experience non-parametrically by storing and retrieving experience through agent scaffolds and test-time compute. In this paper, we show that in environments such as MLE where a good verifier is available, adapting the LM parameters through gradient updates can be more effective in utilizing compute and agent’s experience. Specifically, we show that agents backed by weaker models that improve via reinforcement learning (RL) can eventually outperform agents backed by much larger, but static models for a given MLE task. We identify two major challenges with RL in this setting. First, actions can take a variable amount of time (e.g., executing code for different solutions), which leads to asynchronous policy gradient updates that favor faster but suboptimal solutions. We propose duration-aware gradient updates in a distributed asynchronous RL framework to amplify high-cost but high-reward actions. Second, using performance on the held-out data as a reward for MLE provides limited feedback. A program that’s nearly correct is treated the same as one that fails entirely (e.g., during data loading). We propose environment instrumentation to offer verifiable partial credit, using a separate, static language model to insert print statement to an existing program. Our experiments suggest that a small LM (Qwen2.5-3B) adapted with RL, when given enough compute, can solve an MLE task better than prompting a frontier model (Claude-3.5-Sonnet) with the state-of-the-art agent scaffold (AIDE) by an average of 22% across 12 Kaggle tasks.

Learning expressive and efficient policy functions is a promising direction in reinforcement learning (RL). While flow-based policies have recently proven effective in modeling complex action distributions with a fast deterministic sampling process, they still face a trade-off between expressiveness and computational burden, which is typically controlled by the number of flow steps. In this work, we propose mean flow policy (MFP), a new generative policy function that models the mean velocity field to achieve the fastest one-step action generation. To ensure its high expressiveness, an instantaneous velocity constraint (IVC) is introduced on the mean velocity field during training. We theoretically prove that this design explicitly serves as a crucial boundary condition, thereby improving learning accuracy and enhancing policy expressiveness. Empirically, our MFP achieves state-of-the-art success rates across several challenging robotic manipulation tasks from Robomimic and OGBench. It also delivers substantial improvements in training and inference speed over existing flow-based policy baselines.

Recent works have proposed accelerating the wall-clock training time of actor-critic methods via the use of large-scale environment parallelization; unfortunately, these can sometimes still require large number of environment interactions to achieve a desired level of performance. Noting that well-structured representations can improve the generalization and sample efficiency of deep reinforcement learning (RL) agents, we propose the use of simplicial embeddings: lightweight representation layers that constrain embeddings to simplicial structures. This geometric inductive bias results in sparse and discrete features that stabilize critic bootstrapping and strengthen policy gradients. When applied to FastTD3, FastSAC, and PPO, simplicial embeddings consistently improve sample efficiency and final performance across a variety of continuous- and discrete-control environments, without any loss in runtime speed.

Behavior Foundation Models (BFMs) are capable of retrieving high-performing policy for any reward function specified directly at test-time, commonly referred to as zero-shot reinforcement learning (RL). While this is a very efficient process in terms of compute, it can be less so in terms of data: as a standard assumption, BFMs require computing rewards over a non-negligible inference dataset, assuming either access to a functional form of rewards, or significant labeling efforts. To alleviate these limitations, we tackle the problem of task inference purely through interaction with the environment at test-time. We propose OpTI-BFM, an optimistic decision criterion that directly models uncertainty over reward functions and guides BFMs in data collection for task inference. Formally, we provide a regret bound for well- trained BFMs through a direct connection to upper-confidence algorithms for linear bandits. Empirically, we evaluate OpTI-BFM on established zero-shot benchmarks, and observe that it enables successor-features-based BFMs to identify and optimize an unseen reward function in a handful of episodes with minimal compute overhead.

A hallmark of modern large-scale machine learning techniques is the use of training objectives that provide dense supervision to intermediate computations, such as teacher forcing the next token in language models or denoising step-by-step in diffusion models. This enables models to learn complex functions in a generalizable manner. Motivated by this observation, we investigate the benefits of iterative computation for temporal difference (TD) methods in reinforcement learning (RL). Typically, they represent value functions in a monolithic fashion, without iterative compute. We introduce floq (flow-matching Q-functions), an approach that parameterizes the Q-function using a velocity field and trains it with techniques from flow-matching, typically used in generative modeling. This velocity field underneath the flow is trained using a TD-learning objective, which bootstraps from values produced by a target velocity field, computed by running multiple steps of numerical integration. Crucially, floq allows for more fine-grained control and scaling of the Q-function capacity than monolithic architectures, by appropriately setting the number of integration steps. Across a suite of challenging offline RL benchmarks and online fine-tuning tasks, floq improves performance by nearly 1.8x. floq scales capacity far better than standard TD-learning architectures, highlighting the potential of iterative computation for value learning.

The use of target networks in deep reinforcement learning is a widely popular solution to mitigate the brittleness of semi-gradient approaches and stabilize learning. However, target networks notoriously require additional memory and delay the propagation of Bellman updates compared to an ideal target-free approach. In this work, we step out of the binary choice between target-free and target-based algorithms. We introduce a new method that uses a copy of the last linear layer of the online network as a target network, while sharing the remaining parameters with the up-to-date online network. This simple modification enables us to keep the target-free's low-memory footprint while leveraging the target-based literature. We find that combining our approach with the concept of iterated $Q$-learning, which consists of learning consecutive Bellman updates in parallel, helps improve the sample-efficiency of target-free approaches. Our proposed method, iterated Shared $Q$-Learning (iS-QL), bridges the performance gap between target-free and target-based approaches across various problems while using a single $Q$-network, thus stepping towards resource-efficient reinforcement learning algorithms.