ICLR 2026 Papers

While search-augmented large language models (LLMs) exhibit impressive capabilities, their reliability in complex multi-hop reasoning remains limited. This limitation arises from three fundamental challenges: decomposition errors, where tasks are incorrectly broken down; retrieval missing, where key evidence fails to be retrieved; and reasoning errors, where flawed logic propagates through the reasoning chain. A single failure in any of these stages can derail the final answer. We propose Erasable Reinforcement Learning (ERL), a novel framework that transforms fragile reasoning into a robust process. ERL explicitly identifies faulty steps, erases them, and regenerates reasoning in place, preventing defective logic from propagating through the reasoning chain. This targeted correction mechanism turns brittle reasoning into a more resilient process. Models trained with ERL, termed ESearch, achieve substantial improvements on HotpotQA, MuSiQue, 2Wiki, and Bamboogle, with the 3B model achieving +8.48% EM and +11.56% F1, and the 7B model achieving +5.38% EM and +7.22% F1 over previous state-of-the-art(SOTA) results. These findings suggest that erasable reinforcement learning provides a powerful paradigm shift for robust multi-step reasoning in LLMs.

Unsupervised pre-training can equip reinforcement learning agents with prior knowledge and accelerate learning in downstream tasks. A promising direction, grounded in human development, investigates agents that learn by setting and pursuing their own goals. The core challenge lies in how to effectively generate, select, and learn from such goals. Our focus is on broad distributions of downstream tasks where solving every task zero-shot is infeasible. Such settings naturally arise when the target tasks lie outside of the pre-training distribution or when their identities are unknown to the agent. In this work, we (i) optimize for efficient multi-episode exploration and adaptation within a meta-learning framework, and (ii) guide the training curriculum with evolving estimates of the agent’s post-adaptation performance. We present ULEE, an unsupervised meta-learning method that combines an in-context learner with an adversarial goal-generation strategy that maintains training at the frontier of the agent’s capabilities. On XLand-MiniGrid benchmarks, ULEE pre-training yields improved exploration and adaptation abilities that generalize to novel objectives, environment dynamics, and map structures. The resulting policy attains improved zero-shot and few-shot performance, and provides a strong initialization for longer fine-tuning processes. It outperforms learning from scratch, DIAYN pre-training, and alternative curricula.

Offline reinforcement learning (RL) is a powerful approach for data-driven decision-making and control. Compared to model-free methods, offline model-based reinforcement learning (MBRL) explicitly learns world models from a static dataset and uses them as surrogate simulators, improving the data efficiency and enabling the learned policy to potentially generalize beyond the dataset support. However, there could be various MDPs that behave identically on the offline dataset and dealing with the uncertainty about the true MDP can be challenging. In this paper, we propose modeling offline MBRL as a Bayes Adaptive Markov Decision Process (BAMDP), which is a principled framework for addressing model uncertainty. We further propose a novel Bayes Adaptive Monte-Carlo planning algorithm capable of solving BAMDPs in continuous state and action spaces with stochastic transitions. This planning process is based on Monte Carlo Tree Search and can be integrated into offline MBRL as a policy improvement operator in policy iteration. Our "RL + Search" framework follows in the footsteps of superhuman AIs like AlphaZero, improving on current offline MBRL methods by incorporating more computation input. The proposed algorithm significantly outperforms state-of-the-art offline RL methods on twelve D4RL MuJoCo tasks and three challenging, stochastic tokamak control tasks.

DeepSeek-R1-Zero has successfully demonstrated the emergence of reasoning capabilities in LLMs purely through Reinforcement Learning (RL). Inspired by this breakthrough, we explore how RL can be utilized to enhance the reasoning capability of MLLMs. However, direct training with RL struggles to activate complex reasoning capabilities such as questioning and reflection in MLLMs, due to the absence of substantial high-quality multimodal reasoning data. To address this issue, we propose the reasoning MLLM, Vision-R1, to improve multimodal reasoning capability. Specifically, we first construct a high-quality multimodal CoT dataset without human annotations by leveraging an existing MLLM and DeepSeek-R1 through modality bridging and data filtering to obtain a 200K multimodal CoT dataset, Vision-R1-cold dataset. It serves as cold-start initialization data for Vision-R1. To mitigate the optimization challenges caused by overthinking after cold start, we propose Progressive Thinking Suppression Training (PTST) strategy and employ Group Relative Policy Optimization (GRPO) with the hard formatting result reward function to gradually refine the model's ability to learn correct and complex reasoning processes on the multimodal math dataset. Comprehensive experiments show our model achieves an average improvement of $\sim$6\% across various multimodal math reasoning benchmarks using only a 10K multimodal math data during RL training. Vision-R1-7B achieves a 73.5\% accuracy on the widely used MathVista benchmark, which is only 0.4\% lower than the leading reasoning model, OpenAI O1. Scaling up the amount of multimodal math data in the RL training, Vision-R1-32B and Vison-R1-72B achieves 76.4\% and 78.2\% MathVista benchmark scores, respectively.

Symbolic world modeling is the task of inferring and representing the transitional dynamics of an environment as an executable program. Previous research on symbolic world modeling has focused on simple, deterministic environments with abundant data and human-provided guidance. We address the more realistic and challenging problem of learning a symbolic world model in a complex, stochastic environment with severe constraints: a limited interaction budget where the agent has only “one life” to explore a hostile environment and no external guidance in the form of human-provided, environment-specific rewards or goals. We introduce OneLife, a framework that models world dynamics through conditionally-activated programmatic laws within a probabilistic programming framework. Each law operates through a precondition-effect structure, allowing it to remain silent on irrelevant aspects of the world state and predict only the attributes it directly governs. This creates a dynamic computation graph that routes both inference and optimization only through relevant laws for each transition, avoiding the scaling challenges that arise when all laws must contribute to predictions about a complex, hierarchical state space, and enabling accurate learning of stochastic dynamics even when most rules are inactive at any given moment. To evaluate our approach under these demanding constraints, we introduce a new evaluation protocol that measures (a) state ranking, the ability to distinguish plausible future states from implausible ones, and (b) state fidelity, the ability to generate future states that closely resemble reality. We develop and evaluate our framework on Crafter-OO, our reimplementation of the popular Crafter environment that exposes a structured, object-oriented symbolic state and and a pure transition function that operates on that state alone. OneLife can successfully learn key environment dynamics from minimal, unguided interaction, outperforming a strong baseline on 16 out of 23 scenarios tested. We also demonstrate the world model’s utility for planning, where rollouts simulated within the world model successfully identify superior strategies in multi-step goal-oriented tasks. Our work establishes a foundation for autonomously constructing programmatic world models of unknown, complex environments.

The real world is inherently non-stationary, with ever-changing factors, such as weather conditions and traffic flows, making it challenging for agents to adapt to varying environmental dynamics. Non-Stationary Reinforcement Learning (NSRL) addresses this challenge by training agents to adapt rapidly to sequences of distinct Markov Decision Processes (MDPs). However, existing NSRL approaches often focus on tasks with regularly evolving patterns, leading to limited adaptability in highly dynamic settings. Inspired by the success of Wavelet analysis in time series modeling, specifically its ability to capture signal trends at multiple scales, we propose WISDOM to leverage wavelet-domain predictive task representations to enhance NSRL. WISDOM captures these multi-scale features in evolving MDP sequences by transforming task representation sequences into the wavelet domain, where wavelet coefficients represent both global trends and fine-grained variations of non-stationary changes. In addition to the auto-regressive modeling commonly employed in time series forecasting, we devise a wavelet temporal difference (TD) update operator to enhance tracking and prediction of MDP evolution. We theoretically prove the convergence of this operator and demonstrate policy improvement with wavelet task representations. Experiments on diverse benchmarks show that WISDOM significantly outperforms existing baselines in both sample efficiency and asymptotic performance, demonstrating its remarkable adaptability in complex environments characterized by non-stationary and stochastically evolving tasks.

Exploration remains the key bottleneck for large language model agents trained with reinforcement learning. While prior methods exploit pretrained knowledge, they fail in environments requiring the discovery of novel states. We propose EMPO$^2$, a hybrid RL framework that leverages memory for exploration and combines on- and off-policy updates to make LLMs perform well with memory while also ensuring robustness without it. On ScienceWorld and WebShop, EMPO$^2$ achieves 128.6% and 11.3% improvements over GRPO, respectively. Moreover, in out-of-distribution tests, EMPO$^2$ demonstrates superior adaptability to new tasks, requiring only a few trials with memory and no parameter updates. These results highlight EMPO$^2$ as a promising framework for building more exploratory and generalizable LLM-based agents.

Large Multimodal Models (LMMs) are increasingly capable of answering medical questions that require joint reasoning over images and text, yet training general medical VQA systems is impeded by the lack of large, openly usable, high-quality corpora. We present MedVLSynther, a rubric-guided generator-verifier framework that synthesizes high-quality multiple-choice VQA items directly from open biomedical literature by conditioning on figures, captions, and in-text references. The generator produces self-contained stems and parallel, mutually exclusive options under a machine-checkable JSON schema; a multi-stage verifier enforces essential gates (self-containment, single correct answer, clinical validity, image-text consistency), awards fine-grained positive points, and penalizes common failure modes before acceptance. Applying this pipeline to PubMed Central yields MedSynVQA: 13,087 audited questions over 14,803 images spanning 13 imaging modalities and 28 anatomical regions. Training open-weight LMMs with reinforcement learning using verifiable rewards improves accuracy across six medical VQA benchmarks, achieving averages of 55.85 (3B) and 58.15 (7B), with up to 77.57 on VQA-RAD and 67.76 on PathVQA, outperforming strong medical LMMs. A Ablations verify that both generation and verification are necessary and that more verified data consistently helps, and a targeted contamination analysis detects no leakage from evaluation suites. By operating entirely on open literature and open-weight models, MedVLSynther offers an auditable, reproducible, and privacy-preserving path to scalable medical VQA training data.

Large language models trained with reinforcement learning on verifiable rewards often inflate response length—trading brevity for accuracy. While longer reasoning can help on hard problems, many extra tokens are filler: verbose text making little progress. We introduce GFPO (Group Filtered Policy Optimization), which curbs this length explosion by sampling larger groups per problem and only training on responses filtered by (1) length and (2) token efficiency (reward per token). By sampling more during training time, GFPO teaches models to think less at inference time. On Phi-4-reasoning, GFPO cuts GRPO’s length inflation by up to 85\% across STEM and coding benchmarks (AIME 24/25, GPQA, Omni-MATH, LiveCodeBench) while preserving accuracy. We find that GFPO also outperforms Dr. GRPO in both accuracy and length reduction and generalizes across model sizes and families. We further propose Adaptive Difficulty GFPO, which allocates more training exploration to harder problems, yielding better efficiency-accuracy trade-offs on challenging questions. With only a 7\% increase in training time, GFPO reduces end-to-end latency by $\sim$30\%, cutting response time on hard queries by 90 seconds. GFPO trades modest training-time increases for lasting gains in inference—an effective recipe for efficient reasoning.

The problem of learning how to reach goals in an environment has been a long- standing challenge in for AI researchers. Effective goal-conditioned reinforcement learning (GCRL) methods promise to enable reaching distant goals without task- specific rewards by stitching together past experiences of different complexity. Mathematically, there is a duality between the notion of optimal goal-reaching value functions (the likelihood of success at reaching a goal) and temporal dis- tances (transit times states). Recent works have exploited this property by learning quasimetric distance representations that stitch long-horizon behaviors using the in- ductive bias of their architecture. These methods have shown promise in simulated benchmarks, reducing value learning to a shortest-path problem. But quasimet- ric, and more generally, goal-conditioned RL methods still struggle in complex environments with stochasticity and high-dimensional (visual) observations. There is a fundamental tension between the local dynamic programming (TD backups, temporal distances) that enables optimal shortest-path reasoning in theory and the statistical global MC updates (multistep returns, suboptimal in theory). We show how these approaches can be integrated into a practical GCRL method that fits a quasimetric distance using a multistep Monte-Carlo return. We show our method outperforms existing GCRL methods on long-horizon simulated tasks with up to 4000 steps, even with visual observations. We also demonstrate that our method can enable stitching in the real-world robotic manipulation domain (Bridge setup). Our approach is the first end-to-end GCRL method that enables multistep stitching in this real-world manipulation domain from an unlabeled offline dataset of visual observations.

Data augmentation plays a pivotal role in offline imitation learning (IL) by alleviating covariate shift, yet existing methods remain constrained. Single-step techniques frequently violate underlying system dynamics, whereas trajectory-level approaches are plagued by compounding errors or scalability limitations. Even recent Koopman-based methods typically function at the single-step level, encountering computational bottlenecks due to action-equivariance requirements and vulnerability to approximation errors. To overcome these challenges, we introduce Koopman-Assisted Trajectory Synthesis (KATS), a novel framework for generating complete, multi-step trajectories. By operating at the trajectory level, KATS effectively mitigates compounding errors. It leverages a state-equivariant assumption to ensure computational efficiency and scalability, while incorporating a refined generator matrix to bolster robustness against Koopman approximation errors. This approach enables a more direct and efficacious mechanism for distribution matching in offline IL. Extensive experiments demonstrate that KATS substantially enhances policy performance and achieves state-of-the-art (SOTA) results, especially in demanding scenarios with narrow expert data distributions.

Skill-based reinforcement learning (SBRL) enables rapid adaptation in environments with sparse rewards by pretraining a skill-conditioned policy. Effective skill learning requires jointly maximizing both exploration and skill diversity. However, existing methods often face challenges in simultaneously optimizing for these two conflicting objectives. In this work, we propose a new method, Adaptive Multi-objective Projection for balancing Exploration and skill Diversification (AMPED), which explicitly addresses both: during pre-training, a gradient-surgery projection balances the exploration and diversity gradients, and during fine-tuning, a skill selector exploits the learned diversity by choosing skills suited to downstream tasks. Our approach achieves performance that surpasses SBRL baselines across various benchmarks. Through an extensive ablation study, we identify the role of each component and demonstrate that each element in AMPED is contributing to performance. We further provide theoretical evidence that, with a greedy skill selector, greater skill diversity reduces fine-tuning sample complexity. These results highlight the importance of explicitly harmonizing exploration and diversity and demonstrate the effectiveness of AMPED in enabling robust and generalizable skill learning.

Multi-agent reinforcement learning (MARL) has made significant progress in recent years, but most algorithms still rely on a discrete-time Markov Decision Process (MDP) with fixed decision intervals. This formulation is often ill-suited for complex multi-agent dynamics, particularly in high-frequency or irregular time-interval settings, leading to degraded performance and motivating the development of continuous-time MARL (CT-MARL). Existing CT-MARL methods are mainly built on Hamilton–Jacobi–Bellman (HJB) equations. However, they rarely account for safety constraints such as collision penalties, since these introduce discontinuities that make HJB-based learning difficult. To address this challenge, we propose a continuous-time constrained MDP (CT-CMDP) formulation and a novel MARL framework that transforms discrete MDPs into CT-CMDPs via an epigraph-based reformulation. We then solve this by proposing a novel PINN-based actor–critic method that enables stable and efficient optimization in continuous time. We evaluate our approach on continuous-time safe multi-particle environments (MPE) and safe multi-agent MuJoCo benchmarks. Results demonstrate smoother value approximations, more stable training, and improved performance over safe MARL baselines, validating the effectiveness and robustness of our method.

Powerful large language models (LLMs) from different providers have been expensively trained and finetuned to specialize across varying domains. In this work, we introduce a new kind of Conductor model trained with reinforcement learning to automatically discover powerful coordination strategies among LLMs. Our Conductor learns not only to design targeted communication topologies for effective agent-to-agent collaboration, but also to prompt engineer focused instructions to the LLMs to maximally leverage their individual capabilities. We show that, by learning optimal coordination strategies over pools of powerful worker LLMs, a 7B Conductor achieves significant performance gains beyond any individual worker, attaining state-of-the-art results in challenging reasoning benchmarks, such as LiveCodeBench and GPQA. By training with randomized agent pools, our conductor effectively adapts to arbitrary sets of open- and closed-source agents, meeting any user requirements. Furthermore, allowing the Conductor to select itself as a worker gives rise to recursive topologies, elevating performance with a new form of dynamic test-time scaling through online iterative adaptation. More broadly, ours is among the early work demonstrating language model coordination can be unlocked through RL, where powerful coordination strategies emerge naturally in LLMs through pure end-to-end reward maximization.

Exoskeletons show great promise for enhancing mobility, but providing appropriate assistance remains challenging due to the complexity of human adaptation to external forces. Current state-of-the-art approaches for optimizing exoskeleton controllers require extensive human experiments in which participants must walk for hours, creating a paradox: those who could benefit most from exoskeleton assistance, such as individuals with mobility impairments, are rarely able to participate in such demanding procedures. We present Exo-plore, a simulation framework that combines neuromechanical simulation with deep reinforcement learning to optimize hip exoskeleton assistance without requiring real human experiments. Exo-plore can (1) generate realistic gait data that captures human adaptation to assistive forces, (2) produce reliable optimization results despite the stochastic nature of human gait, and (3) generalize to pathological gaits, showing strong linear relationships between pathology severity and optimal assistance.

In cooperative multi-agent reinforcement learning, agents often face scalability challenges due to the exponential growth of the joint action and observation spaces. Inspired by the structure of human teams, we explore subteam-based coordination, where agents are partitioned into fully correlated subgroups with limited inter-group interaction. We formalize this structure using Bayesian networks and propose a class of correlated joint policies induced by directed acyclic graphs . Theoretically, we prove that regularized policy gradient ascent converges to near-optimal policies under a decomposability condition of the environment. Empirically, we introduce a heuristic for dynamically constructing context-aware subteams with limited dependency budgets, and demonstrate that our method outperforms standard baselines across multiple benchmark environments.

We propose a hierarchical entity-centric framework for offline Goal-Conditioned Reinforcement Learning (GCRL) that combines subgoal decomposition with factored structure to solve long-horizon tasks in domains with multiple entities. Achieving long-horizon goals in complex environments remains a core challenge in Reinforcement Learning (RL). Domains with multiple entities are particularly difficult due to their combinatorial complexity. GCRL facilitates generalization across goals and the use of subgoal structure, but struggles with high-dimensional observations and combinatorial state-spaces, especially under sparse reward. We employ a two-level hierarchy composed of a value-based GCRL agent and a factored subgoal-generating conditional diffusion model. The RL agent and subgoal generator are trained independently and composed post hoc through selective subgoal generation based on the value function, making the approach modular and compatible with existing GCRL algorithms. We introduce new variations to benchmark tasks that highlight the challenges of multi-entity domains, and show that our method consistently boosts performance of the underlying RL agent on image-based long-horizon tasks with sparse rewards, achieving over $150$% higher success rates on the hardest task in our suite and generalizing to increasing horizons and numbers of entities. Rollout videos are provided at: https://sites.google.com/view/hecrl.

We present a novel theoretical framework for deep reinforcement learning (RL) in continuous environments by modeling the problem as a continuous-time stochastic process, drawing on insights from stochastic control. Building on previous work, we introduce a viable model of actor–critic algorithm that incorporates both exploration and stochastic transitions. For single-hidden-layer neural networks, we show that the state of the environment can be formulated as a two time scale process: the environment time and the gradient time. Within this formulation, we characterize how the time-dependent random variables that represent the environment's state and estimate of the cumulative discounted return evolve over gradient steps in the infinite width limit of two-layer networks. Using the theory of stochastic differential equations, we derive, for the first time in continuous RL, an equation describing the infinitesimal change in the state distribution at each gradient step, under a vanishingly small learning rate. Overall, our work provides a novel nonparametric formulation for studying overparametrized neural actor-critic algorithms. We empirically corroborate our theoretical result using a toy continuous control task.

Value decomposition is a core approach for cooperative multi-agent reinforcement learning (MARL). However, existing methods still rely on a single optimal action and struggle to adapt when the underlying value function shifts during training, often converging to suboptimal policies. To address this limitation, we propose Successive Sub-value Q-learning (S2Q), which learns multiple sub-value functions to retain alternative high-value actions. Incorporating these sub-value functions into a Softmax-based behavior policy, S2Q encourages persistent exploration and enables $Q^{\text{tot}}$ to adjust quickly to the changing optima. Experiments on challenging MARL benchmarks confirm that S2Q consistently outperforms various MARL algorithms, demonstrating improved adaptability and overall performance. Our code is available at https://github.com/hyeon1996/S2Q.