ICLR 2026 论文列表

Unsupervised reinforcement learning (URL) pre-trains agents by exploring diverse states in reward-free environments, aiming to enable efficient adaptation to various downstream tasks. Without extrinsic rewards, prior methods rely on intrinsic objectives, but heterogeneous exploration data demand strong modeling capacity for both intrinsic reward design and policy learning. We introduce the **Ex**ploratory **D**iffusion **M**odel (**ExDM**), which leverages the expressive power of diffusion models to fit diverse replay-buffer distributions, thus providing accurate density estimates and a score-based intrinsic reward that drives exploration into under-visited regions. This mechanism substantially broadens state coverage and yields robust pre-trained policies. Beyond exploration, ExDM offers theoretical guarantees and practical algorithms for fine-tuning diffusion policies under limited interactions, overcoming instability and computational overhead from multi-step sampling. Extensive experiments on Maze2d and URLB show that ExDM achieves superior exploration and faster downstream adaptation, establishing new state-of-the-art results, particularly in environments with complex structure or cross-embodiment settings.

While reinforcement learning methods such as Group Relative Preference Optimization (GRPO) have significantly enhanced Large Language Models, adapting them to diffusion models remains challenging. In particular, GRPO demands a stochastic policy, yet the most cost‑effective diffusion samplers are based on deterministic ODEs. Recent work addresses this issue by using inefficient SDE-based samplers to induce stochasticity, but this reliance on model-agnostic Gaussian noise leads to slow convergence. To resolve this conflict, we propose Direct Group Preference Optimization (DGPO), a new online RL algorithm that dispenses with the policy-gradient framework entirely. DGPO learns directly from group-level preferences, which utilize relative information of samples within groups. This design eliminates the need for inefficient stochastic policies, unlocking the use of efficient deterministic ODE samplers and faster training. Extensive results show that DGPO trains around 20 times faster than existing state-of-the-art methods and achieves superior performance on both in-domain and out-of-domain reward metrics.

Does reinforcement learning (RL) teach large language models (LLMs) genuinely new skills, or does it merely activate existing ones? This question lies at the core of ongoing debates about the role of RL in LLM post-training. On one side, strong empirical results can be achieved with RL alone even without preceding supervised finetuning; on the other, critics argue that RL contributes little beyond reweighting existing reasoning strategies. This work provides concrete evidence that LLMs can acquire genuinely new skills during RL by composing existing ones, mirroring one of the central mechanisms by which humans acquire new cognitive skills \citep{Anderson1982Acquisition}. To mitigate data contamination and other confounding factors and to allow precise control over task complexity, we develop a synthetic framework for our investigation. Specifically, we define a skill as the ability to infer the output of a string transformation function $f(x)$ given $x$. Once an LLM has already learned $f$ and $g$ prior to RL, our experiments reveal that RL enables it to learn unseen compositions of them $h(x)=g(f(x))$. Further, this compositional ability generalizes to more difficult problems such as compositions of $>2$ functions unseen during training. Our experiments provide surprising evidence that this compositional ability, acquired on the source task, transfers to a different target task. This transfer occurs even though the model has never trained with RL on any compositional problems in the target task, as long as it has acquired the target task's atomic skills prior to RL on the source task. Our qualitative analysis shows that RL fundamentally changes the reasoning behaviors of the models. In contrast, neither of the findings is observed in next-token prediction training with the same data. Our systematic experiments provide fresh insights into the learning behaviors of widely-used post-training approaches for LLMs. They suggest the value of building base models with the necessary basic skills, followed by RL with appropriate incentivization to acquire more advanced skills that generalize better to complex and out-of-domain problems.

Vision-language alignment in multi-modal large language models (MLLMs) relies on supervised fine-tuning (SFT) or reinforcement learning (RL). To align multi-modal large language models (MLLMs) in the post-training stage, supervised fine-tuning (SFT) is a stable choice but requires human annotations and lacks task generalizations, while Reinforcement Learning (RL) searches for better answers from reward signals but suffers from computational overhead and instability. To achieve balance among scalability, efficiency, and alignment generalizations, we propose MergeMix, a unified paradigm that bridges SFT and RL with an efficient Token Merge based Mixup augmentation. As for the Mixup policy, we generate contextual aligned mixed images with the corresponding labels according to the merged attention maps with cluster regions. Then, we enhance the preference-driven paradigm for MLLMs by building preference pairs with raw images and MergeMix-generated ones and optimizing the soft preference margin with the mixed SimPO loss. Extensive experiments demonstrate that MergeMix not only achieves dominant classification accuracy as an augmentation method but also improves generalization abilities and alignment of MLLMs, providing a new learning paradigm for preference alignment with training efficiency and stability.

Resource allocation (RA) is critical to efficient service deployment in Network Function Virtualization (NFV), a transformative networking paradigm. This task is termed NFV-RA. Recently, deep Reinforcement Learning (RL)-based methods have been showing promising potential to address this combinatorial complexity of constrained cross-graph mapping. However, RL-driven NFV-RA research lacks a systematic benchmark for comprehensive simulation and rigorous evaluation. This gap hinders in-depth performance analysis and slows algorithm development for emerging networks, resulting in fragmented assessments. In this paper, we introduce Virne, a comprehensive benchmarking framework designed to accelerate the research and application of deep RL for NFV-RA. Virne provides customizable simulations for diverse network scenarios, including cloud, edge, and 5G environments. It features a modular and extensible implementation pipeline that integrates over 30 methods of various types. Virne also establishes a rigorous evaluation protocol that extends beyond online effectiveness to include practical perspectives such as solvability, generalizability, and scalability. Furthermore, we conduct in-depth analysis through extensive experiments to provide valuable insights into performance trade-offs for efficient implementation and offer actionable guidance for future research directions. Overall, with its capabilities of diverse simulations, rich implementations, and thorough evaluation, Virne could serve as a comprehensive benchmark for advancing NFV-RA methods and deep RL applications. The code and resources are available at https://anonymous.4open.science/r/anonymous-virne.

Building Behavioral Foundation Models (BFMs) for humanoid robots has the potential to unify diverse control tasks under a single, promptable generalist policy. However, existing approaches are either exclusively deployed on simulated humanoid characters, or specialized to specific tasks such as tracking. We propose BFM-Zero, a framework that learns an effective shared latent representation that embeds motions, goals, and rewards into a common space, enabling a single policy to be prompted for multiple downstream tasks without retraining. This well-structured latent space in BFM-Zero enables versatile and robust whole-body skills on a Unitree G1 humanoid in the real world, via diverse inference methods, including zero-shot motion tracking, goal reaching, and reward inference, and few-shot optimization-based adaptation. Unlike prior on-policy reinforcement learning (RL) frameworks, BFM-Zero builds upon recent advancements in unsupervised RL and Forward-Backward (FB) models, which offer an objective-centric, explainable, and smooth latent representation of whole-body motions. We further extend BFM-Zero with critical reward shaping, domain randomization, and history-dependent asymmetric learning to bridge the sim-to-real gap. Those key design choices are quantitatively ablated in simulation. A first-of-its-kind model, BFM-Zero establishes a step toward scalable, promptable behavioral foundation models for whole-body humanoid control. Videos: http://bfm-zero-anonymous.pages.dev/

Reinforcement learning with verifiable rewards (RLVR) has emerged as a promising paradigm for enhancing the reasoning capabilities of large language models (LLMs). In this context, models explore reasoning trajectories and exploit rollouts with correct answers as positive signals for policy optimization. However, these rollouts might involve flawed patterns such as answer-guessing and jump-in-reasoning. Such flawed-positive rollouts are rewarded identically to fully correct ones, causing policy models to internalize these unreliable reasoning patterns. In this work, we first conduct a systematic study of flawed-positive rollouts in RL and find that they enable rapid capability gains during the early optimization stage, while constraining reasoning capability later by reinforcing unreliable patterns. Building on these insights, we propose **F**lawed-**A**ware **P**olicy **O**ptimization (**FAPO**), which presents a parameter-free reward penalty for flawed-positive rollouts, enabling the policy to leverage them as useful shortcuts in the warm-up stage, securing stable early gains, while gradually shifting optimization toward reliable reasoning in the later refinement stage. To accurately and comprehensively detect flawed-positive rollouts, we introduce a generative reward model (GenRM) with a process-level reward that precisely localizes reasoning errors. Experiments show that FAPO is effective in broad domains, improving outcome correctness, process reliability, and training stability without increasing the token budget.

Behavioral Foundation Models (BFMs) have been recently successful in producing agents with the capabilities to adapt to any unknown reward or task. In reality, these methods are only able to produce near-optimal policies for the reward functions that are in the span of some pre-existing _state features_. Naturally, their efficiency relies heavily on the choice of state features that they use. As a result, these BFMs have used a wide variety of complex objectives, often sensitive to environment coverage, to train task spanning features with different inductive properties. With this work, our aim is to examine the question: are these complex representation learning objectives necessary for zero-shot RL? Specifically, we revisit the objective of self-supervised next-state prediction in latent space for state feature learning, but observe that such an objective alone is prone to increasing state-feature similarity, and subsequently reducing span of reward functions that we can represent optimal policies for. We propose an approach, RLDP, that adds a simple regularization to maintain feature diversity and can match or surpass state-of-the-art complex representation learning methods for zero-shot RL. Furthermore, we demonstrate the prior approaches diverge in low-coverage scenarios where RLDP still succeeds.

Large Language Models (LLMs) have emerged as a promising approach for binary decompilation. However, the existing LLM-based decompilers still are somewhat limited in effectively presenting a program's source-level structure with its original identifiers. To mitigate this, we introduce SK2Decompile, a novel two-phase approach to decompile from the skeleton (semantic structure) to the skin (identifier) of programs. Specifically, we first apply a Structure Recovery model to translate a program's binary code to an Intermediate Representation (IR) as deriving the program's "skeleton", i.e., preserving control flow and data structures while obfuscating all identifiers with generic placeholders. We also apply reinforcement learning to reward the model for producing program structures that adhere to the syntactic and semantic rules expected by compilers. Second, we apply an Identifier Naming model to produce meaningful identifiers which reflect actual program semantics as deriving the program's "skin". We train the Identifier Naming model with a separate reinforcement learning objective that rewards the semantic similarity between its predictions and the reference code. Such a two-phase decompilation process facilitates advancing the correctness and readability of decompilation independently. Our evaluations indicate that SK2Decompile, significantly outperforms the SOTA baselines, achieving 21.6% average re-executability rate gain over GPT-5-mini on the HumanEval dataset and 29.4% average R2I improvement over Idioms on the GitHub2025 benchmark.

Reinforcement learning (RL) has recently become the core paradigm for aligning and strengthening large language models (LLMs). Yet, applying RL in off-policy settings—where stale data from past policies are used for training—improves sample efficiency, but remains challenging: policy entropy declines sharply, optimization often becomes unstable and may even collapse. Through theoretical and empirical analysis, we identify two key insights: (i) an imbalance in optimization, where negative-advantage samples dominate the policy gradient, suppressing useful behaviors and risking gradient explosions; and (ii) the derived Entropy-Clip Rule, which reveals that the fixed clipping mechanism in PPO-like objectives systematically blocks entropy-increasing updates, thereby driving the policy toward over-exploitation at the expense of exploration. Building on these insights, we propose BAlanced Policy Optimization with Adaptive Clipping (BAPO), a simple yet effective method that dynamically adjusts clipping bounds to adaptively re-balance positive and negative contributions, preserve entropy, and stabilize RL optimization. Across diverse off-policy scenarios—including sample replay and partial rollout—BAPO achieves fast, stable, and data-efficient training. On AIME 2024 and AIME 2025 benchmarks, our 7B BAPO model surpasses open-source counterparts such as SkyWork-OR1-7B, while our 32B BAPO model not only achieves state-of-the-art results among models of the same scale but also outperforms leading proprietary systems like o3-mini and Gemini-2.5-Flash-Thinking.

We introduce MedAgentGym, a scalable and interactive training environment designed to enhance coding-based biomedical reasoning capabilities in large language model (LLM) agents. MedAgentGym comprises 72,413 task instances across 129 categories derived from 12 authentic real-world biomedical scenarios. Tasks are encapsulated within executable sandbox environments, each featuring detailed task specifications, interactive feedback mechanisms, verifiable ground truth annotations, and scalable training trajectory generation. Extensive benchmarking of 29 LLMs reveals substantial performance disparities in biomedical data science between commercial and open-source LLMs. Leveraging efficient multi-threaded and multi-turn trajectory sampling in MedAgentGym, Med-Copilot achieves performance gains of +43.02% and +45.28% from offline and online reinforcement learning, respectively, demonstrating MedAgentGym as an effective training ground while establishing itself as a cost-effective, privacy-preserving alternative competitive with proprietary LLMs (gpt-4o). By offering a unified execution environment with a comprehensive benchmark and accessible, extensible training resources, MedAgentGym delivers an integrated platform to develop LLM-based coding assistants for advanced biomedical data science.

Recent advancements in long chain-of-thought (CoT) reasoning, particularly through the Group Relative Policy Optimization algorithm used by DeepSeek-R1, have led to significant interest in the potential of Reinforcement Learning with Verifiable Rewards (RLVR) for Large Language Models (LLMs). While RLVR promises to improve reasoning by allowing models to learn from free exploration, there remains debate over whether it truly enhances reasoning abilities or simply boosts sampling efficiency. This paper systematically investigates the impact of RLVR on LLM reasoning. We revisit Pass@K experiments and demonstrate that RLVR can extend the reasoning boundary for both mathematical and coding tasks. This is supported by our introduction of a novel evaluation metric, CoT-Pass@K, which captures reasoning success by accounting for both the final answer and intermediate reasoning steps. Furthermore, we present a theoretical framework explaining RLVR’s incentive mechanism, demonstrating how it can encourage correct reasoning even when rewards are based solely on answer correctness. Our analysis of RLVR’s training dynamics reveals that it incentivizes correct reasoning early in the process, with substantial improvements in reasoning quality confirmed through extensive evaluations. These findings provide strong evidence of RLVR’s potential to enhance LLM reasoning, offering valuable insights into its mechanisms and performance improvements.

Standard human preference-based alignment methods, such as Reinforcement Learning from Human Feedback (RLHF), are a cornerstone technology for aligning Large Language Models (LLMs) with human values. However, these methods are all underpinned by a strong assumption that the collected preference data is clean and that all observed labels are equally reliable. In reality, large-scale preference datasets contain substantial label noise due to annotator errors, inconsistent instructions, varying expertise, and even adversarial or low-effort feedback. This creates a discrepancy between the recorded data and the ground-truth preferences, which can misguide the model and degrade its performance. To address this challenge, we introduce Robust Preference Optimization (RPO). RPO employs an Expectation-Maximization algorithm to infer the posterior probability of each label’s correctness, which is used to adaptively re-weigh each data point in the training loss to mitigate noise. We further generalize this approach by establishing a theoretical link between arbitrary preference losses and their corresponding probabilistic models. This generalization enables the systematic transformation of existing alignment algorithms into their robust counterparts, elevating RPO from a specific algorithm to a general framework for robust preference alignment. Theoretically, we prove that under the condition of a perfectly calibrated model, RPO is guaranteed to converge to the true noise level of the dataset. Our experiments demonstrate RPO’s effectiveness as a general framework, consistently enhancing four state-of-the-art alignment algorithms (DPO, IPO, SimPO, and CPO). When applied to Mistral and Llama 3 models, the RPO-enhanced methods improve AlpacaEval 2 win rates by up to 7.0 percentage points over their respective baselines.

The Successor Measure (SM), a powerful method in reinforcement learning (RL), describes discounted future state distributions under a policy, and it has recently been studied using generative modeling techniques. Although SM is a powerful predictive object, it lacks compact representations tailored for online RL. To address this, we introduce Successor Flow Features (SF2), a representation learning framework that bridges SM estimation with policy optimization. SF2 leverages flow-matching generative models to approximate successor measures, while enforcing a structured linear decomposition into a time-invariant embedding and a time-dependent projection. This yields compact, policy-aware state-action features that integrate readily into standard off-policy algorithms like TD3 and SAC. Experiments on DeepMind Control Suite tasks show that SF2 improves sample efficiency and training stability compared to strong successor feature baselines. We attribute these gains to the compact representation induced by flow matching, which reduces compounding errors in long-horizon predictions.

Multimodal Large Language Models (MLLMs) have achieved notable gains in various tasks by incorporating Chain-of-Thought (CoT) reasoning in language spaces. Recent work extends this direction by leveraging external tools for visual editing, thereby enhancing the visual signal along the reasoning trajectories. Nevertheless, these approaches remain fundamentally constrained: reasoning is still confined to the language space, with visual information treated as static preconditions. We introduce Latent Visual Reasoning (LVR), a new paradigm that enables autoregressive reasoning directly in the visual embedding space. A visual encoder first projects images into visual tokens within a joint semantic space shared with the language model. The language model is then trained to generate latent states that reconstruct key visual tokens critical for answering the query, constituting the process of latent visual reasoning. By interleaving LVR with standard text generation, our model achieves substantial gains on perception-intensive visual question answering tasks. In addition, we adapt the GRPO algorithm to conduct reinforcement learning on latent reasoning, further balancing LVR and textual generation. We show that LVR substantially improves fine-grained visual understanding and perception, achieving 71.67\% on MMVP compared to 66.67\% with Qwen2.5-VL. Code base and model weights will be released later.

Large-scale pre-training has fundamentally changed how machine learning research is done today: large foundation models are trained once, and then can be used by anyone in the community (including those without data or compute resources to train a model from scratch) to adapt and fine-tune to specific tasks. Applying this same framework to reinforcement learning (RL) is appealing because it offers compelling avenues for addressing core challenges in RL, including sample efficiency and robustness. However, there remains a fundamental challenge to pre-train large models in the context of RL: actions have long-term dependencies, so training a foundation model that reasons across *time* is important. Recent advances in generative AI have provided new tools for modeling highly complex distributions. In this paper, we build a probabilistic model to predict which states an agent will visit in the temporally distant future (i.e., an occupancy measure) using flow matching. As large datasets are often constructed by many distinct users performing distinct tasks, we include in our model a latent variable capturing the user intention. This intention increases the expressivity of our model, and enables adaptation with generalized policy improvement. We call our proposed method **intention-conditioned flow occupancy models (InFOM)**. Comparing with alternative methods for pre-training, our experiments on $36$ state-based and $4$ image-based benchmark tasks demonstrate that the proposed method achieves $1.8 \times$ median improvement in returns and increases success rates by $36\\%$.

Automated theorem proving (ATP) --- the task of generating a proof that passes automated proof verification given a math question in formal language --- is a critical challenge at the intersection of mathematics and Artificial Intelligence (AI). We introduce Goedel-Prover-V2, a family of two language models that establish a new state-of-the-art (SOTA) in open-source ATP, using the Lean proof assistant. In addition to standard expert iteration and reinforcement learning, our approach incorporates three key innovations: (1) During training when improvement plateaus on human questions, the prover does scaffolded data synthesis to generate synthetic questions of increasing difficulty for its own training; (2) The prover is trained to self-correct using Lean compiler feedback; (3) Improved test-time exploration through checkpoint averaging to balance accuracy and diversity. Our small model, Goedel-Prover-V2-8B, reaches 84.6\% pass@32 on MiniF2F and outperforms DeepSeek-Prover-V2-671B despite being $80\times$ smaller. Our flagship model, Goedel-Prover-V2-32B, achieves 88.1\% on MiniF2F at pass@32 in standard mode and 90.4\% in self-correction mode, outperforming prior SOTA by a large margin. Additionally, our flagship model solves 86 problems on PutnamBench at pass@184, securing first place among open-source models and surpassing DeepSeek-Prover-V2-671B's record of 47 problems by pass@1024 with about $20\times$ smaller model size and significantly lower compute budget. Our models, code, and data are released at \url{https://github.com/Goedel-LM/Goedel-Prover-V2}.

Reinforcement Learning with Verifiable Rewards (RLVR) has proven to be a highly effective strategy for empowering Large Language Models (LLMs) with long chain-of-thought reasoning abilities. However, its design and optimizations remain tailored to purely textual domains, resulting in suboptimal performance when applied to multimodal reasoning tasks. In particular, we observe that a major source of error (67%) in current multimodal reasoning lies in the perception of visual inputs. To address this bottleneck, we propose PAPO, a novel policy gradient algorithm that encourages the model to generate visually grounded reasoning without external supervision. Specifically, we introduce the Implicit Perception Loss in the form of a KL divergence term, which maximizes the difference between two probability distributions over the same rollout sequence, conditioned on either the original or corrupted visual input. Notably, PAPO does not rely on any additional data annotation, reward models, or stronger teacher models, and can therefore be seamlessly integrated into mainstream RLVR algorithms such as GRPO and DAPO. To further enhance the training stability of PAPO, we introduce the Double Entropy Loss, which effectively regularizes the new KL objective without compromising performance. Despite its simplicity, PAPO yields significant overall improvements of 4.4%-17.5% on diverse multimodal benchmarks. The improvements are more pronounced, approaching 8.0%-19.1%, on tasks with high vision dependency. We also observe a substantial reduction of 30.5% in perception errors, indicating improved perceptual capabilities with PAPO. Overall, PAPO offers a new perspective on advancing multimodal RLVR via the optimization objective, moving beyond rollout or reward design and pointing toward deeper integration of perception and reasoning.

Reinforcement learning from human feedback (RLHF) has proven effectiveness for aligning text-to-image (T2I) diffusion models with human preferences. Although Direct Preference Optimization (DPO) is widely adopted for its computational efficiency and avoidance of explicit reward modeling, its applications to diffusion models have primarily relied on pairwise preferences. The precise optimization of listwise preferences remains largely unaddressed. In practice, human feedback on image preferences often contains implicit ranked information, which conveys more precise human preferences than pairwise comparisons. In this work, we propose Diffusion-LPO, a simple and effective framework for Listwise Preference Optimization in diffusion models with listwise data. Given a caption, we aggregate user feedback into a ranked list of images and derive a listwise extension of the DPO objective under the Plackett–Luce model. Diffusion-LPO enforces consistency across the entire ranking by encouraging each sample to be preferred over all of its lower-ranked alternatives. We empirically demonstrate the effectiveness of Diffusion-LPO across various tasks, including text-to-image generation, image editing, and personalized preference alignment. Diffusion-LPO consistently outperforms pairwise DPO baselines on visual quality and preference alignment.