ICLR 2026 论文列表

We study the problem of training a risk-sensitive reinforcement learning (RL) agent through imitation learning (IL). Unlike standard IL, our goal is not only to train an agent that matches the expert’s expected return (i.e., its *average performance*) but also its *risk attitude* (i.e., other features of the return distribution, such as variance). We propose a general formulation of the risk-sensitive IL problem in which the objective is to match the expert’s return distribution in Wasserstein distance. We focus on the tabular setting and assume the expert’s reward is *known*. After demonstrating the limited expressivity of Markovian policies for this task, we introduce an efficient and sufficiently expressive subclass of non-Markovian policies tailored to it. Building on this subclass, we develop two provably efficient algorithms—RS-BC and RS-KT —for solving the problem when the transition model is unknown and known, respectively. We show that RS-KT achieves substantially lower sample complexity than RS-BC by exploiting dynamics information. We further demonstrate the sample efficiency of return distribution matching in the setting where the expert’s reward is *unknown* by designing an oracle-based variant of RS-KT. Finally, we complement our theoretical analysis of RS-KT and RS-BC with numerical simulations, highlighting both their sample efficiency and the advantages of non-Markovian policies over standard sample-efficient IL algorithms.

Safe exploration is a prerequisite for deploying reinforcement learning (RL) agents in safety-critical domains. In this paper, we approach safe exploration through the lens of epistemic uncertainty, where the actor’s sensitivity to parameter perturbations serves as a practical proxy for regions of high uncertainty. We propose Sharpness-Aware Policy Optimization (SHAPO), a sharpness-aware policy update rule that evaluates gradients at perturbed parameters, making policy updates pessimistic with respect to the actor’s epistemic uncertainty. Analytically we show that this adjustment implicitly reweighs policy gradients, amplifying the influence of rare unsafe actions while tempering contributions from already safe ones, thereby biasing learning toward conservative behavior in under-explored regions. Across several continuous-control tasks, our method consistently improves both safety and task performance over existing baselines, significantly expanding their Pareto frontiers.

We study off-dynamics offline reinforcement learning, where the goal is to learn a policy from offline source and limited target datasets with mismatched dynamics. Existing methods either penalize the reward or discard source transitions occurring in parts of the transition space with high dynamics shift. As a result, they optimize the policy using data from low-shift regions, limiting exploration of high-reward states in the target domain that do not fall within these regions. Consequently, such methods often fail when the dynamics shift is significant or the optimal trajectories lie outside the low-shift regions. To overcome this limitation, we propose MOBODY, a Model-Based Off-Dynamics Offline RL algorithm that optimizes a policy using learned target dynamics transitions to explore the target domain, rather than only being trained with the low dynamics-shift transitions. For the dynamics learning, built on the observation that achieving the same next state requires taking different actions in different domains, MOBODY employs separate action encoders for each domain to encode different actions to the shared latent space while sharing a unified representation of states and a common transition function. We further introduce a target Q-weighted behavior cloning loss in policy optimization to avoid out-of-distribution actions, which push the policy toward actions with high target-domain Q-values, rather than high source domain Q-values or uniformly imitating all actions in the offline dataset. We evaluate MOBODY on a wide range of MuJoCo and Adroit benchmarks, demonstrating that it outperforms state-of-the-art off-dynamics RL baselines as well as policy learning methods based on different dynamics learning baselines, with especially pronounced improvements in challenging scenarios where existing methods struggle.

Recent advances in reinforcement learning have shown that language models can develop sophisticated reasoning through training on tasks with verifiable rewards, but these approaches depend on human-curated problem-answer pairs and domain-specific reward engineering. We introduce SPIRAL, a self-play framework where models learn by playing multi-turn, zero-sum games against continuously improving versions of themselves, generating an automatic curriculum of stronger opponents, and eliminating the need for human supervision. To enable this self-play training at scale, we implement a fully online, multi-turn, multi-agent reinforcement learning system for LLMs and propose role-conditioned advantage estimation (RAE) to stabilize multi-agent training. SPIRAL produces reasoning capabilities that transfer broadly, improving performance by up to 10\% across a suite of 8 reasoning benchmarks on 4 different models spanning Qwen and Llama model families, outperforming supervised fine-tuning on 25,000 expert game trajectories. Multi-game training (TicTacToe, Kuhn Poker, Simple Negotiation) yields the strongest results, with improvements observed across both base and instruction-tuned models. Analysis of chain-of-thought traces reveals that games develop distinct cognitive patterns that transfer to improve reasoning performance, with different games developing complementary strengths. Even models which have already been trained on reasoning tasks using RLVR, like DeepSeek-R1-Distill-Qwen-7B, still benefit from our approach. These results demonstrate that zero-sum games naturally develop transferable reasoning capabilities across diverse model architectures and training stages, highlighting a promising direction for autonomous reasoning development.

Bayesian Optimization (BO) is a powerful tool for black-box optimization, but its application to high-dimensional permutation spaces is severely limited by the challenge of defining scalable representations. The current state-of-the-art BO approach for permutation spaces relies on an exhaustive $\Omega(n^2)$ pairwise comparison, inducing a dense representation that is impractical for large-scale permutations. To break this barrier, we introduce a novel framework for generating efficient permutation representations via kernel functions derived from sorting algorithms. Within this framework, the Mallows kernel can be viewed as a special instance derived from enumeration sort. Further, we introduce the \textbf{Merge Kernel} , which leverages the divide-and-conquer structure of merge sort to produce a compact, $\Theta(n\log n)$ to achieve the lowest possible complexity with no information loss and effectively capture permutation structure. Our central thesis is that the Merge Kernel performs competitively with the Mallows kernel in low-dimensional settings, but significantly outperforms it in both optimization performance and computational efficiency as the dimension $n$ grows. Extensive evaluations on various permutation optimization benchmarks confirm our hypothesis, demonstrating that the Merge Kernel provides a scalable and more effective solution for Bayesian optimization in high-dimensional permutation spaces, thereby unlocking the potential for tackling previously intractable problems such as large-scale feature ordering and combinatorial neural architecture search.

Existing reinforcement learning (RL) methods struggle with complex dynamical systems that demand interactions at high frequencies or irregular time intervals. Continuous-time RL (CTRL) has emerged as a promising alternative by replacing discrete-time Bellman recursion with differentiable value functions defined as viscosity solutions of the Hamilton–Jacobi–Bellman (HJB) equation. While CTRL has shown promise, its applications have been largely limited to the single-agent domain. This limitation stems from two key challenges: (i) conventional methods for solving HJB equations suffer from the curse of dimensionality (CoD), making them intractable in high-dimensional systems; and (ii) even with learning-based approaches to alleviate the CoD, accurately approximating centralized value functions in multi-agent settings remains difficult, which in turn destabilizes policy training. In this paper, we propose a CT-MARL framework that uses physics-informed neural networks (PINNs) to approximate HJB-based value functions at scale. To ensure the value is consistent with its differential structure, we align value learning with value-gradient learning by introducing a Value Gradient Iteration (VGI) module that iteratively refines value gradients along trajectories. This improves gradient accuracy, in turn yielding more precise value approximations and stronger policy learning. We evaluate our method using continuous‑time variants of standard benchmarks, including multi‑agent particle environment (MPE) and multi‑agent MuJoCo. Our results demonstrate that our approach consistently outperforms existing continuous‑time RL baselines and scales to complex cooperative multi-agent dynamics.

Comparison of fine-tuning models with reinforcement learning (RL) and supervised fine-tuning (SFT) reveals that, despite similar performance at a new task, RL preserves prior knowledge and capabilities significantly better. We find that the degree of forgetting is determined by the distributional shift, measured as the KL-divergence between the fine-tuned and base policy evaluated on the new task. Our analysis reveals that on-policy RL is implicitly biased towards KL-minimal solutions among the many that solve the new task, whereas SFT can converge to distributions arbitrarily far from the base model. We validate these findings through experiments with large language models and robotic foundation models and further provide theoretical justification for why on-policy RL updates lead to a smaller KL change. We term this principle $\textit{RL’s Razor}$: among all ways to solve a new task, RL prefers those closest in KL to the original model.

Active Prompt Learning (APL) using vision-language models (\textit{e.g.}, CLIP) has attracted considerable attention for mitigating the dependence on fully labeled dataset in downstream task adaptation. However, existing methods fail to explicitly leverage prompt to guide sample selection, resulting in the selected samples being ineffective in facilitating the prompt template's downstream task adaptation, while also overlooking valuable complementary information in the unselected samples. To fill this gap, we propose a novel Prompt-Guided Self-Training Sampling Policy (PSP) for APL, which integrates Soft Actor-Critic with a customized real-pseudo hybrid reward and vectorized critics to incorporate prompts in guiding sample selection toward those that facilitate the optimization of prompt template, by jointly considering both selected and unselected samples. Specifically, PSP comprises two prominent components: Vectorized Soft Actor-Critic Sampling Policy (VSSP) and Uncertainty Augmented Self-Training (UST) mechanism. VSSP customizes a real-pseudo hybrid reward based on learned prompts and image features, which is fed into vectorized critics to estimate Q-value for each sample and compute gradients that optimize the actor, allowing it to refine its sampling policy in an End-to-End manner to identify the most informative samples for prompt learning. Moreover, UST leverages the CLIP from the previous round to generate reliable pseudo-labeled data based on uncertainty and confidence of average predictions, thereby deepening the understanding of the overall data. Extensive experiments conducted on diverse real-world datasets validate the effectiveness of our PSP.

Off-policy reinforcement learning (RL) for large language models (LLMs) is attracting growing interest, driven by practical constraints in real-world applications, the complexity of LLM-RL infrastructure, and the need for further innovations of RL methodologies. While classic REINFORCE and its modern variants like Group Relative Policy Optimization (GRPO) are typically regarded as on-policy algorithms with limited tolerance of off-policyness, we present in this work a first-principles derivation for *group-relative REINFORCE* without assuming a specific training data distribution, showing that it admits a *native off-policy interpretation*. This perspective yields two general principles for adapting REINFORCE to truly off-policy settings: regularizing policy updates, and actively shaping the data distribution. Our analysis demystifies some myths about the roles of importance sampling and clipping in GRPO, unifies and reinterprets two recent algorithms --- Online Policy Mirror Descent and Asymmetric REINFORCE --- as regularized forms of the REINFORCE loss, and offers theoretical justification for seemingly heuristic data-weighting strategies. Our findings lead to actionable insights that are validated with extensive empirical studies, and open up new opportunities for principled algorithm design in off-policy RL for LLMs.

We develop a novel memory representation for embodied planning models performing long-horizon mobile manipulation in dynamic, large-scale indoor environments. Prior memory representations fall short in this setting, as they struggle with object movements, suffer from computational deficiencies, and often depend on the heuristic integration of multiple models. To overcome these limitations, we present the Embodied Perception Memory (EMP), a learnable memory designed for embodied planning. EMP is implemented as a unified Vision-Language Model (VLM) that uses egocentric vision to maintain and update a textual environment representation. We further introduce two complementary methods for training planners to leverage the EMP: an imitation strategy that uses human trajectories for natural exploration and interaction, and a novel reinforcement learning approach, Dynamic Difficulty-Aware Fine-Tuning (DDAFT), which improves planning performance via difficulty-aware exploration. Our memory representation, when integrated with our planning training methods, leads to significant improvements on planning tasks, showing up to a 55% increase in success rate on the PARTNR benchmark compared to strong baselines. Also, our planning method outperforms these baselines even when they have access to groundtruth perception.

Reinforcement Learning with Verifiable Reward (RLVR) has equipped large language models (LLMs) with the capability of reasoning over complicated logical problems through policy optimization. However, conventional methods require complete annotation of the entire dataset and allocate computation uniformly over all samples. We articulate the lottery sample hypothesis in policy optimization of LLMs: a large training set contains a small subset that, when trained alone, yields performance comparable to that of the full dataset. This paper therefore explores the following question: How can we identify these lottery-winning samples from the original dataset without access to answers? Unlike prior efforts that analyze the effect of different samples in the training set with complete annotation, this paper focuses on the unsupervised discovery of critical instances for LLM reasoning and proposes a novel framework termed Complementary Conformal Selection (CONST). Specifically, CONST evaluates the importance of samples by considering two complementary components: procedural volatility and outcome volatility. Procedural volatility measures the potential variations during the LLM’s reasoning process, while outcome volatility captures inconsistencies in the final answer. Subsequently, conformal prediction is used to obtain a prediction set whose cardinality serves as the criterion for selecting the lottery-winning samples for annotation. We also provide a theoretical analysis, showing that CONST can effectively approximate the optimal policy. Extensive experiments on various LLMs across different datasets demonstrate the effectiveness of CONST.

Boreal forests store 30-40\% of terrestrial carbon, much in climate-vulnerable permafrost soils, making their management critical for climate mitigation. However, optimizing forest management for both carbon sequestration and permafrost preservation presents complex trade-offs that current tools cannot adequately address. We introduce BoreaRL, the first multi-objective reinforcement learning environment for climate-adaptive boreal forest management, featuring a physically-grounded simulator of coupled energy, carbon, and water fluxes. BoreaRL supports two training paradigms: site-specific mode for controlled studies and generalist mode for learning robust policies under environmental stochasticity. Through evaluation of multi-objective RL algorithms, we reveal a fundamental asymmetry in learning difficulty: carbon objectives are significantly easier to optimize than thaw (permafrost preservation) objectives, with thaw-focused policies showing minimal learning progress across both paradigms. In generalist settings, standard gradient-descent based preference-conditioned approaches fail, while a naive site selection approach achieves superior performance by strategically selecting training episodes. Analysis of learned strategies reveals distinct management philosophies, where carbon-focused policies favor aggressive high-density coniferous stands, while effective multi-objective policies balance species composition and density to protect permafrost while maintaining carbon gains. Our results demonstrate that robust climate-adaptive forest management remains challenging for current MORL methods, establishing BoreaRL as a valuable benchmark for developing more effective approaches. We open-source BoreaRL to accelerate research in multi-objective RL for climate applications.

Reinforcement learning from verifiable rewards (RLVR) is an emerging paradigm for improving the reasoning ability of large language models. However, standard on-policy training discards rollout experiences after a single update, leading to computational inefficiency and instability. While prior work on RL has highlighted the benefits of reusing past experience, the role of experience characteristics in shaping learning dynamics of large reasoning models remains underexplored. In this paper, we are the first to investigate what makes a reasoning experience valuable and identify rollout correctness and entropy as effective indicators of experience value. Based on these insights, we propose ExGRPO (Experiential Group Relative Policy Optimization), a framework that organizes and prioritizes valuable experiences, and employs a mixed-policy objective to balance exploration with experience exploitation. Experiments on five backbone models (1.5B-8B parameters) show that ExGRPO consistently improves reasoning performance on mathematical/general benchmarks, with an average gain of +3.5/7.6 points over on-policy RLVR. Moreover, ExGRPO stabilizes training on both stronger and weaker models where on-policy methods fail. These results highlight principled experience management as a key ingredient for efficient and scalable RLVR.

Generative models have recently demonstrated remarkable success across diverse domains, motivating their adoption as expressive policies in reinforcement learning (RL). While they have shown strong performance in offline RL, particularly where the target distribution is well defined, their extension to online fine-tuning has largely been treated as a direct continuation of offline pre-training, leaving key challenges unaddressed. In this paper, we propose Flow Matching with Injected Noise for Offline-to-Online RL (FINO), a novel method that leverages flow matching-based policies to enhance sample efficiency for offline-to-online RL. FINO facilitates effective exploration by injecting noise into policy training, thereby encouraging a broader range of actions beyond those observed in the offline dataset. In addition to exploration-enhanced flow policy training, we combine an entropy-guided sampling mechanism to balance exploration and exploitation, allowing the policy to adapt its behavior throughout online fine-tuning. Experiments across diverse, challenging tasks demonstrate that FINO consistently achieves superior performance under limited online budgets.

In this paper, we study the distributed experts problem, where $n$ experts are distributed across $s$ servers for $T$ timesteps. The loss of each expert at each time $t$ is the $\ell_p$ norm of the vector that consists of the losses of the expert at each of the $s$ servers at time $t$. The goal is to minimize the regret $R$, i.e., the loss of the distributed protocol compared to the loss of the best expert, amortized over the all $T$ times, while using the minimum amount of communication. We give a protocol that achieves regret roughly $R\gtrsim\frac{1}{\sqrt{T}\cdot\text{poly}\log(nsT)}$, using $\mathcal{O}\left(\frac{n}{R^2}+\frac{s}{R^2}\right)\cdot\max(s^{1-2/p},1)\cdot\text{poly}\log(nsT)$ bits of communication, which improves on previous work.

Despite the significant advances in Deep Reinforcement Learning (RL) observed in the last decade, the amount of training experience necessary to learn effective policies remains one of the primary concerns in both simulated and real environments. Looking to solve this issue, previous work has shown that improved efficiency can be achieved by separately modeling the agent and environment, but usually requires a supervisory signal. In contrast to RL, humans can perfect a new skill from a small number of trials and often do so without a supervisory signal, making neuroscientific studies of human development a valuable source of inspiration for RL. In particular, we explore the idea of motor prediction, which states that humans develop an internal model of themselves and of the consequences that their motor commands have on the immediate sensory inputs. Our insight is that the movement of the agent provides a cue that allows the duality between the agent and environment to be learned. To instantiate this idea, we present Ego-Foresight (EF), a self-supervised method for disentangling agent information based on motion and prediction. Our main finding is that, when used as an auxiliary task in feature learning, self-supervised agent-awareness improves the sample-efficiency and performance of the underlying RL algorithm. To test our approach, we study the ability of EF to predict agent movement and disentangle agent information. Then, we integrate EF with both model-free and model-based RL algorithms to solve simulated control tasks, showing improved sample-efficiency and performance.

Gradient-regularized value learning methods improve sample efficiency by leveraging learned models of transition dynamics and rewards to estimate return gradients. However, existing approaches, such as MAGE, struggle in stochastic or noisy environments, limiting their applicability. In this work, we address these limitations by extending distributional reinforcement learning on continuous state-action spaces to model not only the distribution over scalar state-action value functions but also over their gradients. We refer to this approach as Distributional Sobolev Training. Inspired by Stochastic Value Gradients (SVG), our method utilizes a one-step world model of reward and transition distributions implemented via a conditional Variational Autoencoder (cVAE). The proposed framework is sample-based and employs Max-sliced Maximum Mean Discrepancy (MSMMD) to instantiate the distributional Bellman operator. We prove that the Sobolev-augmented Bellman operator is a contraction with a unique fixed point, and highlight a fundamental smoothness trade-off underlying contraction in gradient-aware RL. To validate our method, we first showcase its effectiveness on a simple stochastic reinforcement‐learning toy problem, then benchmark its performance on several MuJoCo environments.

Multi-turn jailbreaks capture the real threat model for safety-aligned chatbots, where single-turn attacks are merely a special case. Yet existing approaches break under exploration complexity and intent drift. We propose SEMA, a simple yet effective framework that trains a multi-turn attacker without relying on any existing strategies or external data. SEMA comprises two stages. Prefilling self-tuning enables usable rollouts by fine-tuning on non-refusal, well-structured, multi-turn adversarial prompts that are self-generated with a minimal prefix, thereby stabilizing subsequent learning. Reinforcement learning with intent-drift-aware reward trains the attacker to elicit valid multi-turn adversarial prompts while maintaining the same harmful objective. We anchor harmful intent in multi-turn jailbreaks via an intent-drift-aware reward that combines intent alignment, compliance risk, and level of detail. Our open-loop attack regime avoids dependence on victim feedback, unifies single- and multi-turn settings, and reduces exploration complexity. Across multiple datasets, victim models, and jailbreak judges, our method achieves state-of-the-art (SOTA) attack success rates (ASR), outperforming all single-turn baselines, manually scripted and template-driven multi-turn baselines, as well as our SFT (Supervised Fine-Tuning) and DPO (Direct Preference Optimization) variants. For instance, SEMA performs an average 80.1% ASR@1 across three closed-source and open-source victim models on AdvBench, 33.9% over SOTA. The approach is compact, reproducible, and transfers across targets, providing a stronger and more realistic stress test for large language model (LLM) safety and enabling automatic redteaming to expose and localize failure modes.

Gaussian policies have dominated continuous control in deep reinforcement learning (RL), yet they suffer from a fundamental mismatch: their unbounded support requires ad-hoc squashing functions that distort the geometry of bounded action spaces. While von Mises-Fisher (vMF) distributions offer a theoretically grounded alternative on the sphere, their reliance on Bessel functions and rejection sampling hinders practical adoption. We propose \textbf{Geometric Action Control (GAC)}, a novel action generation paradigm that preserves the geometric benefits of spherical distributions while \textit{simplifying computation}. GAC decomposes action generation into a direction vector and a learnable concentration parameter, enabling efficient interpolation between deterministic actions and uniform spherical noise. This design reduces parameter count from \(2d\) to \(d+1\), and avoids the \(O(dk)\) complexity of vMF rejection sampling, achieving simple \(O(d)\) operations. Empirically, GAC consistently matches or exceeds state-of-the-art methods across six MuJoCo benchmarks, achieving 37.6\% improvement over SAC on Ant-v4 and up to 112\% on complex DMControl tasks, demonstrating strong performance across diverse benchmarks. Our ablation studies reveal that both \textbf{spherical normalization} and \textbf{adaptive concentration control} are essential to GAC's success. These findings suggest that robust and efficient continuous control does not require complex distributions, but a principled respect for the geometry of action spaces. Code and pretrained models are available in supplementary materials.