ICLR 2026 Papers

Reinforcement learning (RL) agents achieve remarkable performance but remain far less learning efficient than humans. While RL agents require extensive self-play games to extract useful signals, humans often need only a few games, improving rapidly by repeatedly revisiting states where mistakes occurred. This idea, known as search control, aims to restart from valuable states rather than always from the initial state. In AlphaZero, prior work Go-Exploit applies this idea by sampling past states from self-play or search trees, but it treats all states equally, regardless of their learning potential. We propose Regret-Guided Search Control (RGSC), which extends AlphaZero with a regret network that learns to identify high-regret states, where the agent's evaluation diverges most from the actual outcome. These states are collected from both self-play trajectories and MCTS nodes, stored in a prioritized regret buffer, and reused as new starting positions. Across 9×9 Go, 10×10 Othello, and 11×11 Hex, RGSC outperforms AlphaZero and Go-Exploit by an average of 77 and 89 Elo, respectively. When training on a well-trained 9×9 Go model, RGSC further improves the win rate against KataGo from 69.3% to 78.2%, while both baselines show no improvement. These results demonstrate that RGSC provides an effective mechanism for search control, improving both efficiency and robustness of AlphaZero training.

Data contamination poses a significant threat to the reliable evaluation of Large Language Models (LLMs). This issue arises when benchmark samples may inadvertently appear in training sets, compromising the validity of reported performance. While detection methods have been developed for the pre-training and Supervised Fine-Tuning stages, a critical research gap exists for the increasingly significant phase of Reinforcement Learning (RL) post-training. As RL post-training becomes pivotal for advancing LLM reasoning, the absence of specialized contamination detection methods in this paradigm presents a critical vulnerability. To address this, we conduct the first systematic study of data detection within RL post-training scenario and propose Self-Critique. Our method is motivated by a key observation: after RL phase, the output entropy distribution of LLMs tends to collapse into highly specific and sparse modes. Self-Critique probes for the underlying policy collapse, i.e., the model's convergence to a narrow reasoning path, which causes this entropy reduction. To facilitate this research, we also introduce RL-MIA, a benchmark constructed to simulate this specific contamination scenario. Extensive experiments show that Self-Critique significantly outperforms baseline methods across multiple models and contamination tasks, achieving an AUC improvement of up to 30%. Whereas existing methods are close to a random guess for RL-phase contamination, our method makes detection possible. Our benchmark and code are available at https://anonymous.4open.science/r/Data-Contamination-in-RL.

While offline goal-conditioned reinforcement learning (GCRL) provides a simple recipe to train generalist policies from large unlabeled datasets, Offline GCRL agents trained with sparse rewards typically struggle on long-horizon tasks. Manually designing task-specific reward functions undermines the simplicity, scalability and generality of this paradigm. Moreover, prior approaches to learn rewards for effective credit assignment fail to adequately capture goal-reaching information as tasks scale in complexity. To address this gap, we propose $\textrm{\textbf{Occupancy Reward Shaping(ORS)}}$, a novel reward-shaping approach that leverages a learned occupancy measure; a distribution that naturally captures complex long-horizon temporal dependencies between states; and distills goal-reaching information from the occupancy measure into a general-purpose reward function for effective credit assignment. We demonstrate that ORS achieves a $\mathbf{2.3\times}$ improvement in performance on average over its base RL algorithm across a diverse set of long-horizon locomotion and manipulation tasks and outperforms prior state-of-the-art methods.

When formulating safety as limits of cumulative cost, safe reinforcement learning (RL) learns policies that maximize rewards subject to these constraints during both data collection and deployment. While off-policy methods offer high sample efficiency, their application to safe RL faces substantial challenges from constraint violations caused by the cost-agnostic exploration and the underestimation bias in the cost value function. To address these challenges, we propose Constrained Optimistic eXploration Q-learning (COX-Q), an off-policy primal-dual safe RL method that integrates cost-bounded exploration and conservative distributional RL. First, we introduce a novel cost-constrained optimistic exploration strategy that resolves gradient conflicts between reward and cost in the action space, and adaptively adjusts the trust region to control constraint violation in exploration. Second, we adopt truncated quantile critics to mitigate the underestimation bias in costs. The quantile critics also quantify distributional, risk-sensitive epistemic uncertainty for guiding exploration. Experiments across velocity-constrained robot locomotion, safe navigation, and complex autonomous driving tasks demonstrate that COX-Q achieves high sample efficiency, competitive safety performance during evaluation, and controlled data collection cost in exploration. The results highlight the proposed method as a promising solution for safety-critical RL.

Offline reinforcement learning often relies on behavior regularization that enforces policies to remain close to the dataset distribution. However, such approaches fail to distinguish between high-value and low-value actions in their regularization components. We introduce Guided Flow Policy (GFP), which couples a multi-step flow-matching policy with a distilled one-step actor. The actor directs the flow policy through weighted behavior cloning to focus on cloning high-value actions from the dataset rather than indiscriminately imitating all state-action pairs. In turn, the flow policy constrains the actor to remain aligned with the dataset's best transitions while maximizing the critic. This mutual guidance enables GFP to achieve state-of-the-art performance across 144 state and pixel-based tasks from the OGBench, Minari, and D4RL benchmarks, with substantial gains on suboptimal datasets and challenging tasks.

Policy gradient algorithms have been a driver of much recent advancement in language model reasoning. One of their most appealing properties is the ability to learn from exploration on their own trajectories, a process crucial for discovering diverse approaches and fostering creative solutions. As we show in this paper, most policy gradient algorithms naturally reduce the entropy---and thus the diversity of explored trajectories---as part of training, yielding a policy increasingly limited in its ability to explore. However, not all algorithms exhibit this collapse in entropy equally. In this paper, we formally analyze the contributions of leading policy gradient objectives on entropy, show which mechanisms they employ to implicitly limit entropy collapse, and propose a new regularization method, REPO, that stabilizes entropy over training through the use of an adaptive controller. Models trained with REPO preserve entropy throughout training, yielding final policies that are, on average, more performant. By preserving entropy in the final policy, REPO-trained models can even be re-trained on evolving data distributions in new environments, unlike their non-entropy-preserving counterparts.

Large Reasoning Models (LRMs) demonstrate strong performance in complex tasks but often face the challenge of *overthinking*, leading to substantially high inference costs. Existing approaches synthesize shorter reasoning responses for LRMs to learn, but are inefficient for online usage due to the time-consuming data generation and filtering processes. Meanwhile, online reinforcement learning mainly adopts a length reward to encourage short reasoning responses, but it tends to lose reflection ability and harm performance. To address these issues, we propose REA-RL, which introduces a small reflection model for efficient scaling in online training, offering both parallel sampling and sequential revision. Besides, a reflection reward is designed to further prevent LRMs from favoring short yet non-reflective responses. Experiments show that both methods maintain or enhance performance while significantly improving inference efficiency. Their combination achieves a good balance between performance and efficiency, reducing inference costs by 36\% without compromising performance. Further analysis demonstrates that our methods are effective by maintaining reflection frequency for hard problems while appropriately reducing it for easier ones without losing reflection ability. Code is available at https://anonymous.4open.science/r/REA-RL.

Despite rapid advancements in large language models (LLMs), the token-level autoregressive nature constrains their complex reasoning capabilities. To enhance LLM reasoning, inference-time techniques, including Chain/Tree/Graph-of-Thought(s), successfully improve the performance, as they are fairly cost-effective by guiding reasoning through external logical structures without modifying LLMs' parameters. However, these manually predefined, task-agnostic frameworks are applied uniformly across diverse tasks, lacking adaptability. To improve this, we propose **RL-of-Thoughts (RLoT)**, where we train a lightweight navigator model with reinforcement learning (RL) to generate task-adaptive logical structures at inference time, enhancing LLM reasoning. Specifically, we design five basic logic blocks from the perspective of human cognition. During the reasoning process, the trained RL navigator dynamically selects the suitable logic blocks and combines them into task-specific logical structures according to problem characteristics. Experiments across multiple reasoning benchmarks (AIME, MATH, GPQA, etc.) with multiple LLMs (GPT, Llama, Qwen, and DeepSeek) illustrate that RLoT outperforms established inference-time techniques in most cases and improves up to 13.4% in challenging situations. Remarkably, with less than 3K parameters, our RL navigator is able to make sub-10B LLMs comparable to 100B-scale counterparts. Moreover, the RL navigator demonstrates strong transferability: a model trained on one specific LLM-task pair can effectively generalize to unseen LLMs and tasks. Our code is open-source at https://anonymous.4open.science/r/RL-LLM-Reasoning-1A30.

Reasoning-based image quality assessment (IQA) models trained through reinforcement learning (RL) exhibit exceptional generalization, yet the underlying mechanisms and critical factors driving this capability remain underexplored in current research. Moreover, despite their superior performance, these models incur inference energy usage and latency orders of magnitude higher than their earlier counterparts, restricting their deployment in specific scenarios. Through extensive experiments, this paper verifies and elaborates that through RL training, MLLMs leverage their reasoning capability to convert redundant visual representations into compact, cross-domain aligned text representations. This conversion is precisely the source of the generalization exhibited by these reasoning-based IQA models. Building on this fundamental insight, we propose a novel algorithm, RALI, which employs contrastive learning to directly align images with these generalizable text representations learned by RL. This approach eliminates the reliance on reasoning processes and even obviates the need to load an LLM. For the quality scoring task, this framework achieves generalization performance comparable to reasoning-based models while requiring less than 5% of their model parameters and inference time.

The role of reasoning in Audio Large Language Models remains widely underexplored, as introducing a reasoning process often degrades rather than improves performance during inference, a phenomenon we term test-time inverse scaling, where longer reasoning chains yield progressively worse results. We demonstrate that this stems not from fundamental limitations of reasoning itself, but from inadequate training: models without proper guidance for the reasoning process produce hallucinatory, inconsistent reasoning that accumulates errors over longer chains. To address these challenges, we introduce CESAR (Consistent, Effective, and Scalable Audio Reasoners), shifting from outcome verification to rewarding the reasoning process. Our online reinforcement learning framework employs Group Relative Policy Optimization with a multi-faceted reward suite that incentivizes not only correctness and format but also consistency, structured analytical patterns, causal reasoning, domain-knowledge integration, and calibrated reasoning depth. CESAR resolves test-time inverse scaling, transforming reasoning from detriments into gains while revealing model-specific "reasoning sweet spots", where performance peaks during test-time scaling. We achieve state-of-the-art results on MMAU Test-mini, substantially outperforming Gemini 2.5 Pro and GPT-4o Audio, and near-human-level performance on MMSU reasoning tasks. Through AI-as-judge evaluations and qualitative comparisons, we provide both quantitative and qualitative validation of our improved reasoning quality. Importantly, enhanced reasoning creates synergistic effects, simultaneously improving multimodal reasoning and perception capabilities. Overall, CESAR establishes a principled method for developing robust and scalable reasoning in Audio LLMs.

We propose ERA, a new paradigm for entropy-constrained policy via output activation. It guarantees minimum sampling entropy by transforming the outputs of the last layer. Our approach demonstrates broad effectiveness across different domains: 1) for large language models~(LLMs), boosting the AIME 2025 score for Qwen2.5-Math-7B by 37.4%; 2) for continuous control reinforcement learning agents, improving performance by more than 30% over strong baselines such as SAC on the challenging HumanoidBench; 3) for image classification, enhancing ImageNet top-1 accuracy by 0.69% for ResNet-50. These gains are achieved with a computational overhead of less than 7%. Our work validates output activation as a powerful tool for entropy control, opening a new direction for designing simpler and more robust algorithms.

Replication of experimental results has been a challenge faced by many scientific disciplines, including the field of machine learning. Recent work on the theory of machine learning has formalized replicability as the demand that an algorithm produce identical outcomes when executed twice on different samples from the same distribution. Provably replicable algorithms are especially interesting for reinforcement learning (RL), where algorithms are known to be unstable in practice. While replicable algorithms exist for tabular RL settings, extending these guarantees to more practical function approximation settings has remained an open problem. In this work, we make progress by developing replicable methods for linear function approximation in RL. We first introduce two efficient algorithms for replicable random design regression and uncentered covariance estimation, each of independent interest. We then leverage these tools to provide the first provably efficient replicable RL algorithms for linear Markov decision processes in both the generative model and episodic settings. Finally, we evaluate our algorithms experimentally and show how they can inspire more consistent neural policies.

Reinforcement finetuning (RFT) has garnered significant attention in recent years, particularly for enhancing large reasoning models such as OpenAI o1 and Deepseek R1. The appeal of RFT largely stems from its ability to refine model knowledge, better align outputs with user intent, and address challenges associated with limited finetuning data. Despite these advantages, the application of RFT in large Transformer-based generative agents remains relatively underexplored. Although these agents are designed to address multiple tasks through large-scale autoregressive pretraining and share many properties with large reasoning models, current adaptation strategies predominantly rely on supervised finetuning (SFT). In this work, we conduct a systematic investigation of several RFT techniques across a variety of finetuning parameter configurations and meta-reinforcement learning (meta-RL) environments, employing few-shot offline datasets. We provide a comprehensive analysis of RFT algorithm performance under diverse experimental conditions and, based on our empirical findings, introduce a lightweight enhancement to existing RFT methods. This enhancement consistently improves outcomes by combining the strengths of both SFT and RFT. Our findings provide valuable insights for advancing the effectiveness of RFT approaches and broadening their applicability to meta-RL tasks with large Transformer-based generative agents, motivating further research in broader domains.

Offline multi-agent reinforcement learning (MARL) with multi-task (MT) datasets poses unique challenges, as input structures vary across tasks due to the varying number of agents. Prior works have adopted transformers and hierarchical skill learning to facilitate coordination, but these methods underutilize the transformer’s attention mechanism, focusing instead on extracting transferable skills. Moreover, existing transformer-based approaches compress the entire history into a single token and input this token at next time step, forming simple recursive neural network (RNN) processing on history tokens. As a result, models rely primarily on current and near-past observations while neglecting long historical information, even though the partially observable nature of MARL makes history information critical. In this paper, we propose STAIRS-Former, a transformer architecture augmented with spatial and temporal hierarchies that enables the model to properly attend to critical tokens while effectively leveraging long history. To further enhance robustness across varying token counts, we incorporate token dropout, which improves generalization to diverse agent populations. Experiments on the StarCraft Multi-Agent Challenge (SMAC) benchmark with diverse multi-task datasets show that STAIRS-Former consistently outperforms prior algorithms, achieving new state-of-the-art performance.

Reinforcement learning (RL) research requires diverse, challenging environments that are both tractable and scalable. While modern video games may offer rich dynamics, they are computationally expensive and poorly suited for large-scale experimentation due to their CPU-bound execution. We introduce Octax, a high-performance suite of classic arcade game environments implemented in JAX, based on CHIP-8 emulation, a predecessor to Atari, which is widely adopted as a benchmark in RL research. Octax provides the JAX community with a long-awaited end-to-end GPU alternative to Atari games, offering image-based environments, spanning puzzle, action, and strategy genres, all executable at massive scale on modern GPUs. Our JAX-based implementation achieves orders-of-magnitude speedups over traditional CPU emulators. We demonstrate Octax's capabilities by training RL agents across multiple games, showing significant improvements in training speed and scalability compared to existing solutions. The environment's modular design enables researchers to easily extend the suite with new games or generate novel environments using large language models, making it an ideal platform for large-scale RL experimentation.

Recent work increasingly adopts AI for Science (AI4S) models to replace expensive PDE solvers as simulation environments for reinforcement learning (RL), enabling faster training in complex physical control tasks. However, using approximate simulators introduces modeling errors that affect the learned policy. In this paper, we introduce a unified $\varepsilon$-$N$ framework that quantifies the minimal computational cost $N^*(\varepsilon)$ required for an AI4S model to ensure that tabular RL can estimate the value function with unbiasedness, with probability at least $1 - \delta$. This characterization allows us to connect surrogate accuracy, grid resolution, and RL policy quality under a shared probabilistic language. We analyze how the discretization level $K$ of AI4S and RL space governs both PDE surrogate error and RL lattice approximation error, and we employ spectral theory and Sobolev estimates to derive optimal grid strategies that minimize total cost while preserving learning fidelity. Our theory reveals that different systems, such as ODE- and PDE-governed environments, require different allocations of effort between physical simulation and RL optimization. Overall, our framework offers a principled foundation for designing efficient, scalable, and cost-aware AI4S-RL systems with provable learning guarantees.

Reinforcement learning has emerged as a popular method for post-training large language models (LLMs). While improving the model's performance on downstream tasks, it often reduces the model's output diversity, leading to narrow, canonical responses. Existing methods to enhance diversity are limited, either by operating at inference time or by focusing on lexical differences. We propose a novel training method named DQO (Diversity Quality Optimization) based on determinantal point processes (DPPs) to jointly optimize LLMs for quality and semantic diversity. Our approach samples and embeds a group of responses for each prompt, then uses the determinant of a kernel-based similarity matrix to measure diversity as the volume spanned by the embeddings of these responses. Experiments across instruction-following, summarization, story generation, and reasoning tasks demonstrate that our method substantially improves semantic diversity without sacrificing model quality.

Achieving zero constraint violations in safe reinforcement learning poses a significant challenge. We discover a key obstacle called the safety paradox, where improving policy safety reduces the frequency of constraint-violating samples, thereby impairing feasibility function estimation and ultimately undermining policy safety. We theoretically prove that the estimation error bound of the feasibility function increases as the proportion of violating samples decreases. To overcome the safety paradox, we propose an algorithm called feasible dual policy iteration (FDPI), which employs an additional policy to strategically maximize constraint violations while staying close to the original policy. Samples from both policies are combined for training, with data distribution corrected by importance sampling. Extensive experiments show FDPI's state-of-the-art performance on the Safety-Gymnasium benchmark, achieving the lowest violation and competitive-to-best return simultaneously.

Multi-agent imitation learning (MA-IL) aims to learn optimal policies from expert demonstrations in multi-agent interactive domains. Despite existing guarantees on the performance of the extracted policy, characterizations of its distance to a Nash equilibrium are missing for offline MA-IL. In this paper, we demonstrate impossibility and hardness results of learning low-exploitable policies in general $n$-player Markov Games. We do so by providing examples where even exact measure matching fails, and present challenges associated with the practical case of approximation errors. We then show how these challenges can be overcome using strategic dominance assumptions on the expert equilibrium, assuming BC error $\epsilon_{\text{BC}}$. Specifically, for the case of dominant strategy expert equilibria, this provides a Nash imitation gap of $\mathcal{O}\left(n\epsilon_{\text{BC}}/(1-\gamma)^2\right)$ for a discount factor $\gamma$. We generalize this result with a new notion of best-response continuity, and argue that this is implicitly encouraged by standard regularization techniques.