ICLR 2026 Papers

Reinforcement learning with verifiable rewards (RLVR) has proven effective in training large reasoning models (LRMs) by leveraging answer-verifiable signals to guide policy optimization, which, however, suffers from high annotation costs. To alleviate this problem, recent work has explored unsupervised RLVR methods that derive rewards solely from the model’s internal consistency, such as through entropy and majority voting. While seemingly promising, these methods often suffer from model collapse in the later stages of training, which may arise from the reinforcement of incorrect reasoning patterns in the absence of external supervision. In this work, we investigate a novel semi-supervised RLVR paradigm that utilizes a small labeled set to **guide** RLVR training on unlabeled samples. Our key insight is that supervised rewards are essential for stabilizing consistency-based training on unlabeled samples, ensuring that only reasoning patterns verified on labeled instances are incorporated into RL training. Technically, we propose an effective policy optimization algorithm TraPO that filters out reliable unlabeled samples by matching their learning trajectory similarity to labeled ones. Building on this, TraPO achieves remarkable data efficiency and strong generalization on nine advanced benchmarks. With only 1K labeled and 3K unlabeled samples, TraPO reaches 42.6% average accuracy, surpassing the best unsupervised method trained on 45K unlabeled samples (38.3%). Notably, when using 4K labeled and 12K unlabeled samples, TraPO even **outperforms the fully supervised model** trained on the full 45K labeled samples on all benchmarks, while using only **10%** of the labeled data.

Recent progress in large reasoning models for challenging mathematical reasoning has been driven by reinforcement learning (RL). Incorporating long chain-of-thought (CoT) data during mid-training has also been shown to substantially improve reasoning depth. However, current approaches often utilize CoT data indiscriminately, leaving open the critical question of which data types most effectively enhance model reasoning capabilities. In this paper, we define the foundation model's reasoning potential for the first time as the inverse of the number of independent attempts required to correctly answer the question, which is strongly correlated with the final model performance. We then propose utilizing diverse data enriched with high-value reasoning patterns to expand the reasoning potential. Specifically, we abstract atomic reasoning patterns from CoT sequences, characterized by commonality and inductive capabilities, and use them to construct a core reference set enriched with valuable reasoning patterns. Furthermore, we propose a dual-granularity algorithm involving chains of reasoning patterns and token entropy, efficiently selecting high-value CoT data (CoTP) from the data pool that aligns with the core set, thereby training models to master reasoning effectively. Only 10B-token CoTP data enables the 85A6B Mixture-of-Experts (MoE) model to improve by **9.58\%** on the challenging AIME 2024 and 2025, and to raise the upper bound of downstream RL performance by **7.81\%**.

Recent advances in Large Language Models (LLMs) and Reinforcement Learning (RL) have led to strong performance in open-domain question answering (QA). However, existing models still struggle with questions that admit multiple valid answers. Standard QA benchmarks, which typically assume a single gold answer, overlook this reality and thus produce inappropriate training signals. Existing attempts to handle ambiguity often rely on costly manual annotation, which is difficult to scale to multi-hop datasets such as HotpotQA and MuSiQue. In this paper, we present A$^2$Search, an annotation-free, end-to-end training framework to recognize and handle ambiguity. At its core is an automated pipeline that detects ambiguous questions and gathers alternative answers via trajectory sampling and evidence verification. The model is then optimized with RL using a carefully designed $\mathrm{AnsF1}$ reward, which naturally accommodates multiple answers. Experiments on eight open-domain QA benchmarks demonstrate that A$^2$Search achieves new state-of-the-art performance. With only a single rollout, A$^2$Search-7B yields an average $\mathrm{AnsF1}@1$ score of $48.4$% across four multi-hop benchmarks, outperforming all strong baselines, including the substantially larger ReSearch-32B ($46.2$%). Extensive analyses further show that A$^2$Search resolves ambiguity and generalizes across benchmarks, highlighting that embracing ambiguity is essential for building more reliable QA systems.

Offline–to–online deployment of reinforcement learning (RL) agents often stumbles over two fundamental gaps: (1) the sim-to-real gap, where real-world systems exhibit latency and other physical imperfections not captured in simulation; and (2) the interaction gap, where policies trained purely offline face out-of-distribution (OOD) issues during online execution, as collecting new interaction data is costly or risky. As a result, agents must generalize from static, delay-free datasets to dynamic, delay-prone environments. In this work, we propose $\textbf{DT-CORL}$ ($\textbf{D}$elay-$\textbf{T}$ransformer belief policy $\textbf{C}$onstrained $\textbf{O}$ffline $\textbf{RL}$), a novel framework for learning delay-resilient policies solely from static, delay-free offline data. DT-CORL introduces a transformer-based belief model to infer latent states from delayed observations and jointly trains this belief with a constrained policy objective, ensuring that value estimation and belief representation remain aligned throughout learning. Crucially, our method does not require access to delayed transitions during training and outperforms naive history-augmented baselines, SOTA delayed RL methods, and existing belief-based approaches. Empirically, we demonstrate that DT-CORL achieves strong delay-robust generalization across both locomotion and goal-conditioned tasks in the D4RL benchmark under varying delay regimes. Our results highlight that joint belief-policy optimization is essential for bridging the sim-to-real latency gap and achieving stable performance in delayed environments.

Reasoning LLMs suffer from quadratic compute growth as their context length increases, making reinforcement learning with verifiable rewards (RLVR) and test-time scaling prohibitively expensive. Prior work has tried to lighten the computational burden by shortening reasoning traces through pruning, summarization, or multi-stage training, but these methods remain bound to quadratic costs. We introduce Delethink, a thinking algorithm that realizes the Markovian Thinking Paradigm. Instead of producing one long monolithic reasoning trace, Delethink thinks in a sequence of chunks, the Delethink trace. Each chunk continues reasoning by referring only to a fixed number of prior tokens, which functions as a Markovian state sufficient for progressing reasoning, while deleting the rest. This preserves continuity without carrying the quadratic baggage. As a result, compute scales linearly and peak memory remains constant. In experiments, we show that Delethink can be applied directly to off-the-shelf reasoning models ranging from $1.5\textnormal{B}$ to $30\textnormal{B}$ parameters, with no loss in performance. Extended reasoning becomes possible under fixed memory and linear compute, while enabling efficient RL training on new tasks. On the DeepScaleR dataset, Delethink trains R1DistillQwen1.5B to the same benchmark performance as a standard long chain-of-thought (LongCoT) approach, where both models generate up to $24\textnormal{k}$ thinking tokens. The difference is efficiency. Delethink reasons $40\%$ faster with $70\%$ less memory footprint. By decoupling reasoning length from context length, the Markovian Thinking paradigm opens the door to next-generation reasoning LLMs that can scale to millions of tokens with linear compute and constant memory.

Reinforcement Learning (RL) has been shown to significantly boost reasoning capabilities of large language models (LLMs) in math, coding, and multi-hop reasoning tasks. However, RL fine-tuning requires abundant high-quality verifiable data, often obtained through human-annotated datasets and LLM-as-verifier loops. Both of these data types have considerable limitations: human-annotated datasets are small and expensive to curate, while LLM verifiers have high scoring latency and are costly to operate. In this work, we investigate the use of synthetic datasets in RL fine-tuning for multi-hop reasoning tasks. We discover that LLMs fine-tuned on synthetic data perform significantly better on popular real-world question-answering benchmarks, even though the synthetic data only contain fictional knowledge. On stratifying model performance by question difficulty, we find that synthetic data teaches LLMs to compose knowledge, which we to be a fundamental and generalizable reasoning skill. Our work thus highlights the utility of synthetic reasoning datasets in improving LLM reasoning capabilities.

Recent reinforcement learning (RL) methods have substantially enhanced the planning capabilities of Large Language Models (LLMs), yet the theoretical basis for their effectiveness remains elusive. In this work, we investigate RL's benefits and limitations through a tractable graph-based abstraction, focusing on policy gradient (PG) and Q-learning methods. Our theoretical analyses reveal that supervised fine-tuning (SFT) may introduce co-occurrence-based spurious solutions, whereas RL achieves correct planning primarily through exploration, underscoring exploration’s role in enabling better generalization. However, we also show that PG suffers from diversity collapse, where output diversity decreases during training and persists even after perfect accuracy is attained. By contrast, Q-learning provides two key advantages: off-policy learning and diversity preservation at convergence. We further demonstrate that careful reward design is necessary to prevent Q-value bias in Q-learning. Finally, applying our framework to the real-world planning benchmark Blocksworld, we confirm that these behaviors manifest in practice.

Generative models have gained significant traction in offline reinforcement learning (RL) due to their ability to model complex trajectory distributions. However, existing generation-based approaches still struggle with long-horizon tasks characterized by sparse rewards. Some hierarchical generation methods have been developed to mitigate this issue by decomposing the original problem into shorter-horizon subproblems using one policy and generating detailed actions with another. While effective, these methods often overlook the multi-scale temporal structure inherent in trajectories, resulting in suboptimal performance. To overcome these limitations, we propose MAGE, a Multi-scale Autoregressive GEneration-based offline RL method. MAGE incorporates a condition-guided multi-scale autoencoder to learn hierarchical trajectory representations, along with a multi-scale transformer that autoregressively generates trajectory representations from coarse to fine temporal scales. MAGE effectively captures temporal dependencies of trajectories at multiple resolutions. Additionally, a condition-guided decoder is employed to exert precise control over short-term behaviors. Extensive experiments on five offline RL benchmarks against fifteen baseline algorithms show that MAGE successfully integrates multi-scale trajectory modeling with conditional guidance, generating coherent and controllable trajectories in long-horizon sparse-reward settings.

Proximal Policy Optimization (PPO)-based reinforcement learning from human feedback (RLHF) is a widely adopted paradigm for aligning large language models (LLMs) with human preferences. However, its training pipeline suffers from substantial inefficiencies due to sequential multi-model dependencies (e.g., reward model depends on actor outputs) and long-tail response lengths, where a few long responses straggle the stage completion. We present OPPO, a novel, lightweight, and model-agnostic PPO-based RLHF framework that improves training efficiency by overlapping pipeline execution. OPPO introduces two novel techniques: (1) Intra-step overlap, which streams upstream model outputs (e.g., actor model) in right-sized chunks, enabling the downstream model (e.g., reward) to begin prefill while the upstream continues decoding; and (2) Inter-step overlap, which adaptively overcommits a few prompts and defers long generations to future steps, mitigating tail latency without discarding partial work. OPPO integrates easily with existing PPO implementations with a lightweight wrapper. Extensive evaluations show that OPPO accelerates PPO-based RLHF training by $1.8\times$--$2.8\times$ and improves GPU utilization by $1.4\times$--$2.1\times$ without compromising training convergence.

Large Reasoning Models (LRMs) have achieved impressive performance on challenging tasks, yet their deep reasoning often incurs substantial computational costs. To achieve efficient reasoning, existing reinforcement learning methods still struggle to construct short reasoning path during the rollout stage, limiting effective learning. Inspired by Evidence Accumulation Models, we find that LRMs have accumulated sufficient information early in reasoning, making further reasoning steps redundant. Based on this insight, we propose Just-Enough Thinking (JET), which trains models to proactively terminate unnecessary reasoning. JET performs trajectory truncation during rollout to expose the model to short, distributionally consistent reasoning paths. Besides, it uses a quality-controlled length reward to better encourage concise reasoning while maintaining correctness. Extensive experiments demonstrate that JET significantly improves reasoning efficiency without sacrificing accuracy. In particular, JET delivers a 4.6% accuracy improvement while reducing the output length by 46.3% on the Olympiad benchmark using DeepSeek-R1-Distill-Qwen-1.5B. Our code is available in the GitHub.

Learning in high-dimensional MDPs with complex state dynamics became possible with the progress achieved in reinforcement learning research. At the same time, deep neural policies have been observed to be highly unstable with respect to the minor variations in their state space, causing volatile and unpredictable behaviour. To alleviate these volatilities, a line of work suggested techniques to cope with this problem via explicitly regularizing the temporal difference loss to ensure local $\epsilon$-invariance in the state space. In this paper, we provide theoretical foundations on the impact of $\epsilon$-local invariance training on the deep neural policy manifolds. Our comprehensive theoretical and experimental analysis reveals that standard reinforcement learning inherently learns counterfactual values while recent training techniques that focus on explicitly enforcing $\epsilon$-local invariance cause policies to lose counterfactuality, and further result in learning misaligned and inconsistent values. In connection to this analysis, we further highlight that this line of training methods break the core intuition and the true biological inspiration of reinforcement learning, and introduce an intrinsic gap between how natural intelligence understands and interacts with an environment in contrast to AI agents trained via $\epsilon$-local invariance methods. The misalignment, inaccuracy and the loss of counterfactuality revealed in our paper further demonstrate the need to rethink the approach in establishing truly reliable and generalizable reinforcement learning policies.

Few-shot learning (FSL) aims to generalize to novel categories with only a few samples. Recent approaches incorporate large language models (LLMs) to enrich visual representations with semantic embeddings derived from class names. However, they overlook progressive and adaptive alignment between vision and language from low-level to high-level semantics, resulting in limited semantic gains. To address these challenges, we propose Dual-level Vision–Language Alignment with Reinforcement Learning gating (DVLA-RL), which consists of Dual-level Semantic Construction (DSC) and RL-gated Attention (RLA). Specifically, DSC conditions LLMs on both class names and support samples to generate discriminative attributes, progressively selects the most relevant ones, and then synthesizes them into coherent class descriptions. This process provides complementary low-level attributes and high-level descriptions, enabling both fine-grained grounding and holistic class understanding. To dynamically integrate dual-level semantics along with the visual network layers, RLA formulates cross-modal fusion as a sequential decision process. A lightweight policy trained with episodic REINFORCE adaptively adjusts the contributions of self-attention and cross-attention to integrate textual and visual tokens. As a result, shallow layers refine local attributes and deep layers emphasize global semantics, enabling more precise cross-modal alignment. This achieves class-specific discrimination and generalized representations with merely a few support samples. DVLA-RL achieves new state-of-the-art performance across nine benchmarks in three diverse FSL scenarios.

The remarkable success of multimodal large language models (MLLMs) has driven advances in multimodal embeddings, yet existing models remain inherently discriminative, limiting their ability to benefit from reasoning-driven generation paradigm. In this work, we pioneer the exploration of generative embeddings, unifying embedding tasks within a generative paradigm. We propose UME-R1, a universal multimodal embedding framework consisting of a two-stage training strategy: a cold-start supervised fine-tuning equips the model with reasoning capabilities and enables it to generate both discriminative and generative embeddings; a subsequent reinforcement learning enhances reasoning and further optimizes generative embedding quality. This pioneering work reveals four key insights: 1) generative embeddings unlock substantial performance gains over conventional discriminative embeddings by leveraging the powerful generative reasoning capabilities of MLLMs; 2) discriminative and generative embeddings are complementary, whose combined oracle performance far exceeding that of either alone; 3) RL can effectively enhance generative embeddings, establishing a scalable optimization paradigm.; 4) repeated sampling at inference boosts downstream task coverage (pass@k), highlighting the inference-time scalability potential of generative embeddings. Evaluated on the MMEB-V2 benchmark across 78 tasks spanning video, image, and visual documents, UME-R1 significantly outperforms conventional discriminative embedding models and offers a foundation for more interpretable, reasoning-driven generative multimodal embeddings.

Effective information seeking in multi-turn medical dialogues is critical for accurate diagnosis, especially when dealing with incomplete information. Aligning Large Language Models (LLMs) for these interactive scenarios is challenging due to the uncertainty inherent in user-agent interactions, which we formulate as a Hierarchical Markov Decision Process (H-MDP). While conventional Reinforcement Learning (RL) methods like Group Relative Policy Optimization (GRPO) struggle with long-horizon credit assignment and Proximal Policy Optimization (PPO) suffers from unstable value estimation in this context, we propose a novel uncertainty-aware Adaptive Tree Policy Optimization (ATPO) algorithm. Our method adaptively allocates the rollout budget to states with high uncertainty, quantified by a composite metric of Bellman error and action-value variance. This strategy enables more accurate value estimation, while fostering more efficient and diverse exploration. To mitigate the high computational cost of tree-based RL, we introduce two key optimizations: an uncertainty-guided pruning mechanism to minimize the number of rollouts, and an asynchronous search architecture that leverages KV cache reuse to maximize inference throughput. Extensive experiments on three public medical dialogue benchmarks demonstrate that our algorithm significantly outperforms several strong baselines, culminating in Qwen3-8B model surpassing the much larger GPT-4o (+0.92% accuracy).

Reinforcement learning for safety-critical tasks requires policies that are both high-performing and safe throughout the learning process. While model-predictive shielding is a promising approach, existing methods are often computationally intractable for the high-dimensional, nonlinear systems where deep RL excels, as they typically rely on a patchwork of local models. We introduce **RAMPS**, a scalable shielding framework that overcomes this limitation by leveraging a learned, linear representation of the environment's dynamics. This model can range from a linear regression in the original state space to a more complex operator learned in a high-dimensional feature space. The key is that this linear structure enables a robust, look-ahead safety technique based on a *multi-step Control Barrier Function (CBF)*. By moving beyond myopic one-step formulations, **RAMPS** accounts for model error and control delays to provide reliable, real-time interventions. The resulting framework is minimally invasive, computationally efficient, and built upon robust control-theoretic foundations. Our experiments demonstrate that **RAMPS** significantly reduces safety violations compared to existing safe RL methods while maintaining high task performance in complex control environments.

Medical Large Vision-Language Models (Med-LVLMs) have shown strong potential in multimodal diagnostic tasks. However, existing single-agent models struggle to generalize across diverse medical specialties, limiting their performance. Recent efforts introduce multi-agent collaboration frameworks inspired by clinical workflows, where general practitioners (GPs) and specialists interact in a fixed sequence. Despite improvements, these static pipelines lack flexibility and adaptability in reasoning. To address this, we propose MMedAgent-RL, a reinforcement learning (RL)-based multi-agent framework that enables dynamic, optimized collaboration among medical agents. Specifically, we train two GP agents based on Qwen2.5-VL via RL: the triage doctor learns to assign patients to appropriate specialties, while the attending physician integrates the judgments from multi-specialists and its own knowledge to make final decisions. To address the inconsistency in specialist outputs, we introduce a curriculum learning (CL)-guided RL strategy with dynamic entropy regulation, progressively teaching the attending physician to balance between imitating specialists and correcting their mistakes. Experiments on five medical VQA benchmarks demonstrate that MMedAgent-RL outperforms both open-source and proprietary Med-LVLMs. Notably, it achieves an average performance gain of 23.6\% over strong baselines.

In this work, we contribute the first approach to solve infinite-horizon discounted general-utility Markov decision processes (GUMDPs) in the single-trial regime, i.e., when the agent's performance is evaluated based on a single trajectory. First, we provide some fundamental results regarding policy optimization in the single-trial regime, investigating which class of policies suffices for optimality, casting our problem as a particular MDP that is equivalent to our original problem, as well as studying the computational hardness of policy optimization in the single-trial regime. Second, we show how we can leverage online planning techniques, in particular a Monte-Carlo tree search algorithm, to solve GUMDPs in the single-trial regime. Third, we provide experimental results showcasing the superior performance of our approach in comparison to relevant baselines.

Large language models (LLMs), despite possessing latent safety understanding from their vast pretraining data, remain vulnerable to generating harmful content and exhibit issues such as over-refusal and utility degradation after safety alignment. Current safety alignment methods often result in superficial refusal shortcuts or rely on intensive supervision for reasoning-based approaches, failing to fully leverage the model's intrinsic safety self-awareness. We propose \textbf{AlphaAlign}, a simple yet effective pure reinforcement learning (RL) framework with verifiable safety reward designed to incentivize this latent safety awareness through proactive safety reasoning. AlphaAlign employs a dual-reward system: a verifiable safety reward encourages correctly formatted and explicitly justified refusals for harmful queries while penalizing over-refusals, and a normalized helpfulness reward guides high-quality responses to benign inputs. This allows the model to develop proactive safety reasoning capabilities without depending on supervised safety-specific reasoning data. AlphaAlign demonstrates three key advantages: (1) Simplicity and efficiency, requiring only binary prompt safety labels and minimal RL steps for substantial improvements. (2) Breaking the safety-utility trade-off, by enhancing refusal of harmful content and reducing over-refusals, while simultaneously maintaining or even improving general task performance and robustness to unseen jailbreaks. (3) Deep alignment, fostering proactive safety reasoning that generates explicit safety rationales rather than relying on shallow refusal patterns. Our codes are available at \url{https://anonymous.4open.science/r/AlphaAlign-D5B0}

While most reinforcement learning methods today flatten the distribution of future returns to a single scalar value, distributional RL methods exploit the return distribution to provide stronger learning signals and to enable applications in exploration and safe RL. While the predominant method for estimating the return distribution is by modeling it as a categorical distribution over discrete bins or estimating a finite number of quantiles, such approaches leave unanswered questions about the fine-grained structure of the return distribution and about how to distinguish states with high return uncertainty for decision-making. The key idea in this paper is to use modern, flexible flow-based models to estimate the full future return distributions and identify those states with high return variance. We do so by formulating a new flow-matching objective that generates probability density paths satisfying the distributional Bellman equation. Building upon the learned flow models, we estimate the return uncertainty of distinct states using a new flow derivative ODE. We additionally use this uncertainty information to prioritize learning a more accurate return estimation on certain transitions. We compare our method (Value Flows) with prior methods in the offline and online-to-online settings. Experiments on $37$ state-based and $25$ image-based benchmark tasks demonstrate that Value Flows achieves a $1.3\times$ improvement on average in success rates.