ICLR 2026 Papers

Reasoning models have recently shown remarkable progress in domains such as math and coding. However, their expert-level abilities in math and coding contrast sharply with their performance in long-horizon, interactive tasks such as web navigation and computer/phone-use. Inspired by literature on human cognition, we argue that current AI agents need ``vicarious trial and error'' - the capacity to mentally simulate alternative futures before acting - in order to enhance their understanding and performance in complex interactive environments. We introduce Dyna-Mind, a two-stage training framework that explicitly teaches (V)LM agents to integrate such simulation into their reasoning. In stage 1, we introduce Reasoning with Simulations (ReSim), which trains the agent to generate structured reasoning traces from expanded search trees built from real experience gathered through environment interactions. ReSim thus grounds the agent's reasoning in faithful world dynamics and equips it with the ability to anticipate future states in its reasoning. In stage 2, we propose Dyna-GRPO, an online reinforcement learning method to further strengthen the agent's simulation and decision-making ability by using both outcome rewards and intermediate states as feedback from real rollouts. Experiments on two synthetic benchmarks (Sokoban and ALFWorld) and one realistic benchmark (AndroidWorld) demonstrate that (1) ReSim effectively infuses simulation ability into AI agents, and (2) Dyna-GRPO leverages outcome and interaction-level signals to learn better policies for long-horizon, planning-intensive tasks. Together, these results highlight the central role of simulation in enabling AI agents to reason, plan, and act more effectively in the ever more challenging environments.

This study proposes FlexiVoice, a text-to-speech (TTS) synthesis system capable of flexible style control with zero-shot voice cloning. The speaking style is controlled by a natural-language instruction and the voice timbre is provided by a speech reference in zero-shot manner. FlexiVoice is built with an LLM core, which takes text as input, and also takes an optional natural language instruction and an optional speech reference to control style and timbre, respectively. FlexiVoice is equipped with a novel Progressive Post-Training (PPT) scheme that progressively unlocks accurate and flexible controllability. In particular, it first employs Direct Preference Optimization (DPO) to enable FlexiVoice to accurately follow both natural language instruction and speech reference simultaneously. It then uses a multi-objective Group Relative Policy Optimization (GRPO) to disentangle style instruction, reference timbre, and textual content. Finally, it adapts instruction GRPO for more advanced instruction following. Experimental results show that FlexiVoice surpasses competing baselines and demonstrates strong capability in decoupling control factors. Human evaluations further confirm its naturalness, controllability, and robustness. Audio samples are available at https://flexi-voice.github.io/.

The rapid advancement of Multimodal Large Language Models (MLLMs) has made aligning them with human preferences a critical challenge. Reward Models (RMs) are a core technology for achieving this goal, but a systematic guide for building state-of-the-art Multimodal Reward Models (MRMs) is currently lacking in both academia and industry. Through exhaustive experimental analysis, this paper aims to provide a clear “recipe” for constructing high-performance MRMs. We systematically investigate every crucial component in the MRM development pipeline, including reward modeling paradigms (e.g., Naive-RM, Critic-based RM, and Generative RM), reward head architecture, training strategies, data curation (covering over ten multimodal and text-only preference datasets), backbone model and model scale, and ensemble methods. Based on these experimental insights, we introduce BaseReward, a powerful and efficient baseline for multimodal reward modeling. BaseReward adopts a simple yet effective architecture, built upon a Qwen2.5-VL backbone, featuring an optimized two-layer reward head, and is trained on a carefully curated mixture of high-quality multimodal and text-only preference data. Our results show that BaseReward establishes a new state-of-the-art (SOTA) on major benchmarks such as MM-RLHF-Reward Bench, VL-Reward Bench, and Multimodal Reward Bench, outperforming previous open-source and proprietary models. Furthermore, to validate its practical utility beyond static benchmarks, we integrate BaseReward into a real-world reinforcement learning pipeline, successfully enhancing an MLLM’s performance across various perception, reasoning, and conversational tasks. This work not only delivers a top-tier MRM but, more importantly, provides the community with a clear, empirically backed guide for developing robust reward models for the next generation of MLLMs.

Addressing climate change requires global coordination, yet rational economic actors often prioritize immediate gains over collective welfare, resulting in social dilemmas. InvestESG is a recently proposed multi-agent simulation that captures the dynamic interplay between investors and companies under climate risk. We provide a formal characterization of the conditions under which InvestESG exhibits an intertemporal social dilemma, deriving theoretical thresholds at which individual incentives diverge from collective welfare. Building on this, we apply Advantage Alignment, a scalable opponent shaping algorithm shown to be effective in general-sum games, to influence agent learning in InvestESG. We offer theoretical insights into why Advantage Alignment systematically favors socially beneficial equilibria by biasing learning dynamics toward cooperative outcomes. Our results demonstrate that strategically shaping the learning processes of economic agents can result in better outcomes that could inform policy mechanisms to better align market incentives with long-term sustainability goals.

Large Language Models (LLMs) can enhance their reasoning by interacting with external tools, a paradigm known as Tool-Integrated Reasoning (TIR). However, extending TIR to multi-turn settings using Reinforcement Learning (RL) often exhibits training instability and degraded performance. We attribute the instability to harmful negative samples resulting from distributional drift and compounding errors induced by using external tool outputs during multi-turn rollout. To address this issue, we introduce SimpleTIR, a simple method that stabilizes multi-turn TIR training via filtering out trajectories with "void turns", i.e., turns that yield neither a code block nor a final answer. Specifically, we remove those trajectories from the policy update to block harmful gradients, while retaining them in advantage estimation to keep the estimate unbiased. Extensive experiments show that SimpleTIR effectively mitigates gradient norm explosion and stabilizes multi-turn RL training from base models. It achieves state-of-the-art performance on challenging math reasoning benchmarks, including an AIME24 score of 50.5 starting from the Qwen2.5-7B base model. SimpleTIR also promotes more diverse reasoning behaviors such as self-correction and cross-validation, outperforming prior methods trained from stronger instruction-tuned models.

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.

Despite recent rapid progress in AI safety, current large language models remain vulnerable to adversarial attacks in multi-turn interaction settings, where attackers strategically adapt their prompts across conversation turns and pose a more critical yet realistic challenge. Existing approaches that discover safety vulnerabilities either rely on manual red-teaming with human experts or employ automated methods using pre-defined templates and human-curated attack data, with most focusing on single-turn attacks. However, these methods did not explore the vast space of possible multi-turn attacks, failing to consider novel attack trajectories that emerge from complex dialogue dynamics and strategic conversation planning. This gap is particularly critical given recent findings that LLMs exhibit significantly higher vulnerability to multi-turn attacks compared to single-turn attacks. We propose DialTree-RPO, an on-policy reinforcement learning framework integrated with tree search that autonomously discovers diverse multi-turn attack strategies by treating the dialogue as a sequential decision-making problem, enabling systematic exploration without manually curated data. Through extensive experiments, our approach not only achieves more than 25.9% higher ASR across 10 target models compared to previous state-of-the-art approaches, but also effectively uncovers new attack strategies by learning optimal dialogue policies that maximize attack success across multiple turns.

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.

Conventional preference learning methods often prioritize opinions held more widely when aggregating preferences from multiple evaluators. This may result in policies that are biased in favor of some types of opinions or groups and susceptible to strategic manipulation. To address this issue, we develop a novel preference learning framework capable of aligning aggregate opinions and policies proportionally with the true population distribution of evaluator preferences. Grounded in social choice theory, our approach infers the feasible set of evaluator population distributions directly from pairwise comparison data. Using these estimates, the algorithm constructs a policy that satisfies foundational axioms from social choice theory, namely monotonicity and Pareto efficiency, as well as our newly-introduced axioms of population-proportional alignment and population-bounded manipulability. Moreover, we propose a soft-max relaxation method that smoothly trade-offs population-proportional alignment with the selection of the Condorcet winner (which beats all other options in pairwise comparisons). Finally, we validate the effectiveness and scalability of our approach through experiments on both tabular recommendation tasks and large language model alignment.

We present MetaSpatial, the first reinforcement learning (RL) framework for enhancing 3D spatial reasoning in vision-language models (VLMs), enabling real-time 3D scene layout generation without post-processing. MetaSpatial addresses two key challenges: (i) the need for extensive post-processing, as existing VLMs lack inherent 3D spatial reasoning to generate realistic layouts; and (ii) the inefficiency of supervised fine-tuning (SFT) for layout generation due to scarcity of perfect annotations. Our core contribution is the 3D Spatial Policy Optimization (3D-SPO) algorithm, which incorporates physics-aware modulation into advantage estimates at the object level and trajectory-level reward from a training-only multi-turn refinement pipeline. This design enhances temporal credit assignment and encourages spatially consistent policy learning. Empirical evaluations across models of varying scales demonstrate that MetaSpatial improves spatial coherence, physical plausibility, and formatting stability, leading to more realistic and functionally coherent object placements applicable to metaverse environments.

Inference-time compute scaling has emerged as a powerful paradigm for improving language model performance on a wide range of tasks, but the question of how best to use the additional compute remains open. A popular approach is *Best-of-$N$* (BoN) sampling, where $N$ candidate responses are generated, scored according to a reward model, and the highest-scoring response is selected. While this approach can improve performance, it is vulnerable to *reward hacking*, where performance degrades as $N$ increases due to the selection of responses that exploit imperfections in the reward model instead of genuinely improving generation quality. Prior attempts to mitigate reward hacking---via stronger reward models or heavy-handed distributional regularization---either fail to fully address over-optimization or are too conservative to exploit additional compute. In this work, we explore the principle of *pessimism* in reinforcement learning (RL), which uses lower confidence bounds on value estimates to avoid out-of-distribution (OOD) actions with uncertain reward estimates. Our approach, termed as *caution*, can be seen as the *reverse* of *curiosity*: where curiosity (e.g., via Random Network Distillation, RND) rewards prediction error as a signal of novelty, caution penalizes prediction error as a signal of distributional uncertainty. Practically, caution trains an error model on typical responses and uses its prediction error to lower reward estimates for atypical ones. Our extensive empirical evaluation demonstrates that caution is a simple, computationally efficient approach that substantially mitigates reward hacking in BoN sampling. We also provide a theoretical analysis in a simplified linear setting, which shows that caution provably improves over the standard BoN approach. Together, our results not only establish caution as a practical solution to reward hacking, but also provide evidence that curiosity-based approaches can be a general OOD detection technique in LLM settings.

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.

Large language models (LLMs) often generate hallucinations---unsupported content that undermines reliability. While most prior works frame hallucination detection as a binary task, many real-world applications require identifying hallucinated spans, which is a multi-step decision making process. This naturally raises the question of whether explicit reasoning can help the complex task of detecting hallucination spans. To answer this question, we first evaluate pretrained models with and without Chain-of-Thought (CoT) reasoning, and show that CoT reasoning has the potential to generate at least one correct answer when sampled multiple times. Motivated by this, we propose RL4HS, a reinforcement learning framework that incentivizes reasoning with a span-level reward function. RL4HS builds on Group Relative Policy Optimization and introduces Class-Aware Policy Optimization to mitigate reward imbalance issue. Experiments on the RAGTruth benchmark (summarization, question answering, data-to-text) show that RL4HS surpasses pretrained reasoning models and supervised fine-tuning, demonstrating the necessity of reinforcement learning with span-level rewards for detecting hallucination spans.

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.

We present MARTI (Multi-Agent Reinforced Training and Inference), an open-source framework designed to facilitate scalable and efficient learning of multi-agent LLM systems. MARTI supports centralized multi-agent interactions and distributed policy training, with the added capability of multi-turn asynchronous rollouts to enhance training efficiency. The framework includes dynamic workflows for multi-agent interactions, which integrate both rule-based verifiable rewards and LLM-based generative rewards. We validate the effectiveness of MARTI through comprehensive experiments on diverse mathematical tasks, demonstrating that multi-agent LLM-based systems outperform single-agent systems within the same inference budget after convergence. Our contributions lay the foundation for exploring scalable collaborations within LLM-based multi-agent systems and advancing the capabilities of large reasoning models.

Instruction-guided image editing has achieved remarkable progress, yet current models still face challenges with complex instructions and often require multiple samples to produce a desired result. Reinforcement Learning (RL) offers a promising solution, but its adoption in image editing has been severely hindered by the lack of a high-fidelity, efficient reward signal. In this work, we present a comprehensive methodology to overcome this barrier, centered on the development of a state-of-the-art, specialized reward model. We first introduce $\textbf{EditReward-Bench}$, a comprehensive benchmark to systematically evaluate reward models on editing quality. Guided by this benchmark, we develop $\textbf{EditScore}$, an efficient model to evaluate the quality of instruction-guided editing. Through meticulous data curation and filtering, EditScore effectively matches the performance of learning proprietary VLMs. Furthermore, coupled with an effective self-ensemble strategy tailored for the generative nature of EditScore, our largest variant even surpasses GPT-5 in the benchmark. We then demonstrate that a high-fidelity reward model is the key to unlocking online RL for image editing. Our experiments show that, while even the largest open-source VLMs fail to provide an effective learning signal, EditScore enables efficient and robust policy optimization. Applying our framework to a strong base model, OmniGen2, results in a final model that shows a substantial and consistent performance uplift. Overall, this work provides the first systematic path from benchmarking to reward modeling to RL training in image editing, showing that a high-fidelity, domain-specialized reward model is the key to unlocking the full potential of RL in this domain. Our code, models, and benchmark will be released publicly.

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.