Existing reasoning segmentation approaches typically fine-tune multimodal large language models (MLLMs) using image-text pairs and corresponding mask labels. However, they exhibit limited generalization to out-of-distribution scenarios without an explicit reasoning process. Although recent efforts leverage reinforcement learning through group-relative policy optimization (GRPO) to enhance reasoning ability, they often suffer from overthinking - producing uniformly verbose reasoning chains irrespective of task complexity. This results in elevated computational costs and limited control over reasoning quality. To address this problem, we propose PixelThink, a simple yet effective scheme that integrates externally estimated task difficulty and internally measured model uncertainty to regulate reasoning generation within a reinforcement learning paradigm. The model learns to compress reasoning length in accordance with scene complexity and predictive confidence. To support comprehensive evaluation, we introduce ReasonSeg-Diff, an extended benchmark with annotated reasoning references and difficulty scores, along with a suite of metrics designed to assess segmentation accuracy, reasoning quality, and efficiency jointly. Experimental results demonstrate that the proposed approach improves both reasoning efficiency and overall segmentation performance. Our work contributes novel perspectives towards efficient and interpretable multimodal understanding. The code and model will be publicly available.
Coordinating the motion between lower and upper limbs and aligning limb control with perception are substantial challenges in robotics, particularly in dynamic environments. To this end, we introduce an approach for enabling legged mobile manipulators to play badminton, a task that requires precise coordination of perception, locomotion, and arm swinging. We propose a unified reinforcement learning-based control policy for whole-body visuomotor skills involving all degrees of freedom to achieve effective shuttlecock tracking and striking. This policy is informed by a perception noise model that utilizes real-world camera data, allowing for consistent perception error levels between simulation and deployment and encouraging learned active perception behaviors. Our method includes a shuttlecock prediction model, constrained reinforcement learning for robust motion control, and integrated system identification techniques to enhance deployment readiness. Extensive experimental results in a variety of environments validate the robot's capability to predict shuttlecock trajectories, navigate the service area effectively, and execute precise strikes against human players, demonstrating the feasibility of using legged mobile manipulators in complex and dynamic sports scenarios.
Large language models excel at complex tasks by breaking down problems into structured reasoning steps. However, reasoning traces often extend beyond reaching a correct answer, causing wasted computation, reduced readability, and hallucinations. To address this, we introduce a novel hyperparameter-free conciseness score used as a reward signal within a reinforcement learning framework to guide models toward generating correct and concise reasoning traces. This score is evaluated by a large language model acting as a judge, enabling dynamic, context-aware feedback beyond simple token length. Our method achieves state-of-the-art efficiency-accuracy trade-offs on the MATH dataset, reducing token usage by up to 31x on simple problems while improving accuracy by 7%, and on the hardest problems, it outperforms full reasoning by +7.5% accuracy with up to 3.6x fewer tokens. On TheoremQA, our method improves accuracy by +2.2% using 12.5x fewer tokens. We also conduct ablation studies on the judge model, reward composition, and problem difficulty, showing that our method dynamically adapts reasoning length based on problem difficulty and benefits significantly from stronger judges. The code, model weights, and datasets are open-sourced at https://github.com/RazvanDu/ConciseRL.
In the era of Industry 4.0 and smart manufacturing, process systems engineering must adapt to digital transformation. While reinforcement learning offers a model-free approach to process control, its applications are limited by the dependence on accurate digital twins and well-designed reward functions. To address these limitations, this paper introduces a novel framework that integrates inverse reinforcement learning (IRL) with multi-task learning for data-driven, multi-mode control design. Using historical closed-loop data as expert demonstrations, IRL extracts optimal reward functions and control policies. A latent-context variable is incorporated to distinguish modes, enabling the training of mode-specific controllers. Case studies on a continuous stirred tank reactor and a fed-batch bioreactor validate the effectiveness of this framework in handling multi-mode data and training adaptable controllers.
In multi-agent reinforcement learning (MARL), effective communication improves agent performance, particularly under partial observability. We propose MARL-CPC, a framework that enables communication among fully decentralized, independent agents without parameter sharing. MARL-CPC incorporates a message learning model based on collective predictive coding (CPC) from emergent communication research. Unlike conventional methods that treat messages as part of the action space and assume cooperation, MARL-CPC links messages to state inference, supporting communication in non-cooperative, reward-independent settings. We introduce two algorithms -Bandit-CPC and IPPO-CPC- and evaluate them in non-cooperative MARL tasks. Benchmarks show that both outperform standard message-as-action approaches, establishing effective communication even when messages offer no direct benefit to the sender. These results highlight MARL-CPC's potential for enabling coordination in complex, decentralized environments.
Sometimes we benefit from actions that others have taken even when we are unaware that they took those actions. For example, if your neighbor chooses not to take a parking spot in front of your house when you are not there, you can benefit, even without being aware that they took this action. These "hidden gifts" represent an interesting challenge for multi-agent reinforcement learning (MARL), since assigning credit when the beneficial actions of others are hidden is non-trivial. Here, we study the impact of hidden gifts with a very simple MARL task. In this task, agents in a grid-world environment have individual doors to unlock in order to obtain individual rewards. As well, if all the agents unlock their door the group receives a larger collective reward. However, there is only one key for all of the doors, such that the collective reward can only be obtained when the agents drop the key for others after they use it. Notably, there is nothing to indicate to an agent that the other agents have dropped the key, thus the act of dropping the key for others is a "hidden gift". We show that several different state-of-the-art RL algorithms, including MARL algorithms, fail to learn how to obtain the collective reward in this simple task. Interestingly, we find that independent model-free policy gradient agents can solve the task when we provide them with information about their own action history, but MARL agents still cannot solve the task with action history. Finally, we derive a correction term for these independent agents, inspired by learning aware approaches, which reduces the variance in learning and helps them to converge to collective success more reliably. These results show that credit assignment in multi-agent settings can be particularly challenging in the presence of "hidden gifts", and demonstrate that learning awareness in independent agents can benefit these settings.
Recent advances have demonstrated the effectiveness of Reinforcement Learning (RL) in improving the reasoning capabilities of Large Language Models (LLMs). However, existing works inevitably rely on high-quality instructions and verifiable rewards for effective training, both of which are often difficult to obtain in specialized domains. In this paper, we propose Self-play Reinforcement Learning(SeRL) to bootstrap LLM training with limited initial data. Specifically, SeRL comprises two complementary modules: self-instruction and self-rewarding. The former module generates additional instructions based on the available data at each training step, employing robust online filtering strategies to ensure instruction quality, diversity, and difficulty. The latter module introduces a simple yet effective majority-voting mechanism to estimate response rewards for additional instructions, eliminating the need for external annotations. Finally, SeRL performs conventional RL based on the generated data, facilitating iterative self-play learning. Extensive experiments on various reasoning benchmarks and across different LLM backbones demonstrate that the proposed SeRL yields results superior to its counterparts and achieves performance on par with those obtained by high-quality data with verifiable rewards. Our code is available at https://github.com/wantbook-book/SeRL.
Reward models trained with conventional Reinforcement Learning from AI Feedback (RLAIF) methods suffer from limited generalizability, which hinders the alignment performance of the policy model during reinforcement learning (RL). This challenge stems from various issues, including distribution shift, preference label noise, and mismatches between overly challenging samples and model capacity. In this paper, we attempt to enhance the generalizability of reward models through a data-centric approach, driven by the insight that these issues are inherently intertwined from the perspective of data difficulty. To address this, we propose a novel framework, $\textit{Curriculum-RLAIF}$, which constructs preference pairs with varying difficulty levels and produces a curriculum that progressively incorporates preference pairs of increasing difficulty for reward model training. Our experimental results suggest that reward models trained with Curriculum-RLAIF achieve improved generalizability, significantly increasing the alignment performance of the policy model by a large margin without incurring additional inference costs compared to various non-curriculum baselines. Detailed analysis and comparisons with alternative approaches, including data selection via external pretrained reward models or internal self-selection mechanisms, as well as other curriculum strategies, further demonstrate the superiority of our approach in terms of simplicity, efficiency, and effectiveness.
Recent advances in reinforcement learning with verifiable, rule-based rewards have greatly enhanced the reasoning capabilities and out-of-distribution generalization of VLMs/LLMs, obviating the need for manually crafted reasoning chains. Despite these promising developments in the general domain, their translation to medical imaging remains limited. Current medical reinforcement fine-tuning (RFT) methods predominantly focus on close-ended VQA, thereby restricting the model's ability to engage in world knowledge retrieval and flexible task adaptation. More critically, these methods fall short of addressing the critical clinical demand for open-ended, reasoning-intensive decision-making. To bridge this gap, we introduce \textbf{MedCCO}, the first multimodal reinforcement learning framework tailored for medical VQA that unifies close-ended and open-ended data within a curriculum-driven RFT paradigm. Specifically, MedCCO is initially fine-tuned on a diverse set of close-ended medical VQA tasks to establish domain-grounded reasoning capabilities, and is then progressively adapted to open-ended tasks to foster deeper knowledge enhancement and clinical interpretability. We validate MedCCO across eight challenging medical VQA benchmarks, spanning both close-ended and open-ended settings. Experimental results show that MedCCO consistently enhances performance and generalization, achieving a 11.4\% accuracy gain across three in-domain tasks, and a 5.7\% improvement on five out-of-domain benchmarks. These findings highlight the promise of curriculum-guided RL in advancing robust, clinically-relevant reasoning in medical multimodal language models.
Reinforcement learning with outcome-based feedback faces a fundamental challenge: when rewards are only observed at trajectory endpoints, how do we assign credit to the right actions? This paper provides the first comprehensive analysis of this problem in online RL with general function approximation. We develop a provably sample-efficient algorithm achieving $\widetilde{O}({C_{\rm cov} H^3}/{\epsilon^2})$ sample complexity, where $C_{\rm cov}$ is the coverability coefficient of the underlying MDP. By leveraging general function approximation, our approach works effectively in large or infinite state spaces where tabular methods fail, requiring only that value functions and reward functions can be represented by appropriate function classes. Our results also characterize when outcome-based feedback is statistically separated from per-step rewards, revealing an unavoidable exponential separation for certain MDPs. For deterministic MDPs, we show how to eliminate the completeness assumption, dramatically simplifying the algorithm. We further extend our approach to preference-based feedback settings, proving that equivalent statistical efficiency can be achieved even under more limited information. Together, these results constitute a theoretical foundation for understanding the statistical properties of outcome-based reinforcement learning.
Reinforcement learning with verifiable rewards (RLVR) has boosted math and coding in large language models, yet there has been little effort to extend RLVR into messier, real-world domains like forecasting. One sticking point is that outcome-based reinforcement learning for forecasting must learn from binary, delayed, and noisy rewards, a regime where standard fine-tuning is brittle. We show that outcome-only online RL on a 14B model can match frontier-scale accuracy and surpass it in calibration and hypothetical prediction market betting by adapting two leading algorithms, Group-Relative Policy Optimisation (GRPO) and ReMax, to the forecasting setting. Our adaptations remove per-question variance scaling in GRPO, apply baseline-subtracted advantages in ReMax, hydrate training with 100k temporally consistent synthetic questions, and introduce lightweight guard-rails that penalise gibberish, non-English responses and missing rationales, enabling a single stable pass over 110k events. Scaling ReMax to 110k questions and ensembling seven predictions yields a 14B model that matches frontier baseline o1 on accuracy on our holdout set (Brier = 0.193, p = 0.23) while beating it in calibration (ECE = 0.042, p < 0.001). A simple trading rule turns this calibration edge into \$127 of hypothetical profit versus \$92 for o1 (p = 0.037). This demonstrates that refined RLVR methods can convert small-scale LLMs into potentially economically valuable forecasting tools, with implications for scaling this to larger models.
Reinforcement learning (RL) is increasingly applied to real-world problems involving complex and structured decisions, such as routing, scheduling, and assortment planning. These settings challenge standard RL algorithms, which struggle to scale, generalize, and exploit structure in the presence of combinatorial action spaces. We propose Structured Reinforcement Learning (SRL), a novel actor-critic framework that embeds combinatorial optimization layers into the actor neural network. We enable end-to-end learning of the actor via Fenchel-Young losses and provide a geometric interpretation of SRL as a primal-dual algorithm in the dual of the moment polytope. Across six environments with exogenous and endogenous uncertainty, SRL matches or surpasses the performance of unstructured RL and imitation learning on static tasks and improves over these baselines by up to 92% on dynamic problems, with improved stability and convergence speed.
Offline reinforcement learning (RL) aims to learn an optimal policy from pre-collected data. However, it faces challenges of distributional shift, where the learned policy may encounter unseen scenarios not covered in the offline data. Additionally, numerous applications suffer from a scarcity of labeled reward data. Relying on labeled data alone often leads to a narrow state-action distribution, further amplifying the distributional shift, and resulting in suboptimal policy learning. To address these issues, we first recognize that the volume of unlabeled data is typically substantially larger than that of labeled data. We then propose a semi-pessimistic RL method to effectively leverage abundant unlabeled data. Our approach offers several advantages. It considerably simplifies the learning process, as it seeks a lower bound of the reward function, rather than that of the Q-function or state transition function. It is highly flexible, and can be integrated with a range of model-free and model-based RL algorithms. It enjoys the guaranteed improvement when utilizing vast unlabeled data, but requires much less restrictive conditions. We compare our method with a number of alternative solutions, both analytically and numerically, and demonstrate its clear competitiveness. We further illustrate with an application to adaptive deep brain stimulation for Parkinson's disease.
Reinforcement fine-tuning (RFT) has emerged as a powerful post-training technique to incentivize the reasoning ability of large language models (LLMs). However, LLMs can respond very inconsistently to RFT: some show substantial performance gains, while others plateau or even degrade. To understand this divergence, we analyze the per-step influence of the RL objective and identify two key conditions for effective post-training: (1) RL-informative rollout accuracy, and (2) strong data co-influence, which quantifies how much the training data affects performance on other samples. Guided by these insights, we propose behavior injection, a task-agnostic data-augmentation scheme applied prior to RL. Behavior injection enriches the supervised finetuning (SFT) data by seeding exploratory and exploitative behaviors, effectively making the model more RL-ready. We evaluate our method across two reasoning benchmarks with multiple base models. The results demonstrate that our theoretically motivated augmentation can significantly increases the performance gain from RFT over the pre-RL model.
Recent advances in reinforcement learning (RL) have demonstrated the powerful exploration capabilities and multimodality of generative diffusion-based policies. While substantial progress has been made in offline RL and off-policy RL settings, integrating diffusion policies into on-policy frameworks like PPO remains underexplored. This gap is particularly significant given the widespread use of large-scale parallel GPU-accelerated simulators, such as IsaacLab, which are optimized for on-policy RL algorithms and enable rapid training of complex robotic tasks. A key challenge lies in computing state-action log-likelihoods under diffusion policies, which is straightforward for Gaussian policies but intractable for flow-based models due to irreversible forward-reverse processes and discretization errors (e.g., Euler-Maruyama approximations). To bridge this gap, we propose GenPO, a generative policy optimization framework that leverages exact diffusion inversion to construct invertible action mappings. GenPO introduces a novel doubled dummy action mechanism that enables invertibility via alternating updates, resolving log-likelihood computation barriers. Furthermore, we also use the action log-likelihood for unbiased entropy and KL divergence estimation, enabling KL-adaptive learning rates and entropy regularization in on-policy updates. Extensive experiments on eight IsaacLab benchmarks, including legged locomotion (Ant, Humanoid, Anymal-D, Unitree H1, Go2), dexterous manipulation (Shadow Hand), aerial control (Quadcopter), and robotic arm tasks (Franka), demonstrate GenPO's superiority over existing RL baselines. Notably, GenPO is the first method to successfully integrate diffusion policies into on-policy RL, unlocking their potential for large-scale parallelized training and real-world robotic deployment.
Although Large Language Models (LLMs) have demonstrated remarkable progress, their proficiency in graph-related tasks remains notably limited, hindering the development of truly general-purpose models. Previous attempts, including pretraining graph foundation models or employing supervised fine-tuning, often face challenges such as the scarcity of large-scale, universally represented graph data. We introduce G1, a simple yet effective approach demonstrating that Reinforcement Learning (RL) on synthetic graph-theoretic tasks can significantly scale LLMs' graph reasoning abilities. To enable RL training, we curate Erd\~os, the largest graph reasoning dataset to date comprising 50 diverse graph-theoretic tasks of varying difficulty levels, 100k training data and 5k test data, all drived from real-world graphs. With RL on Erd\~os, G1 obtains substantial improvements in graph reasoning, where our finetuned 3B model even outperforms Qwen2.5-72B-Instruct (24x size). RL-trained models also show strong zero-shot generalization to unseen tasks, domains, and graph encoding schemes, including other graph-theoretic benchmarks as well as real-world node classification and link prediction tasks, without compromising general reasoning abilities. Our findings offer an efficient, scalable path for building strong graph reasoners by finetuning LLMs with RL on graph-theoretic tasks, which combines the strengths of pretrained LLM capabilities with abundant, automatically generated synthetic data, suggesting that LLMs possess graph understanding abilities that RL can elicit successfully.
Transfer reinforcement learning aims to derive a near-optimal policy for a target environment with limited data by leveraging abundant data from related source domains. However, it faces two key challenges: the lack of performance guarantees for the transferred policy, which can lead to undesired actions, and the risk of negative transfer when multiple source domains are involved. We propose a novel framework based on the pessimism principle, which constructs and optimizes a conservative estimation of the target domain's performance. Our framework effectively addresses the two challenges by providing an optimized lower bound on target performance, ensuring safe and reliable decisions, and by exhibiting monotonic improvement with respect to the quality of the source domains, thereby avoiding negative transfer. We construct two types of conservative estimations, rigorously characterize their effectiveness, and develop efficient distributed algorithms with convergence guarantees. Our framework provides a theoretically sound and practically robust solution for transfer learning in reinforcement learning.
Offline reinforcement learning (RL) recovers the optimal policy $\pi$ given historical observations of an agent. In practice, $\pi$ is modeled as a weighted version of the agent's behavior policy $\mu$, using a weight function $w$ working as a critic of the agent's behavior. Though recent approaches to offline RL based on diffusion models have exhibited promising results, the computation of the required scores is challenging due to their dependence on the unknown $w$. In this work, we alleviate this issue by constructing a diffusion over both the actions and the weights. With the proposed setting, the required scores are directly obtained from the diffusion model without learning extra networks. Our main conceptual contribution is a novel guidance method, where guidance (which is a function of $w$) comes from the same diffusion model, therefore, our proposal is termed Self-Weighted Guidance (SWG). We show that SWG generates samples from the desired distribution on toy examples and performs on par with state-of-the-art methods on D4RL's challenging environments, while maintaining a streamlined training pipeline. We further validate SWG through ablation studies on weight formulations and scalability.
In this paper, we propose a robust and reinforcement-learning-enhanced network intrusion detection system (NIDS) designed for class-imbalanced and few-shot attack scenarios in Industrial Internet of Things (IIoT) environments. Our model integrates a TabTransformer for effective tabular feature representation with Proximal Policy Optimization (PPO) to optimize classification decisions via policy learning. Evaluated on the TON\textunderscore IoT benchmark, our method achieves a macro F1-score of 97.73\% and accuracy of 98.85\%. Remarkably, even on extremely rare classes like man-in-the-middle (MITM), our model achieves an F1-score of 88.79\%, showcasing strong robustness and few-shot detection capabilities. Extensive ablation experiments confirm the complementary roles of TabTransformer and PPO in mitigating class imbalance and improving generalization. These results highlight the potential of combining transformer-based tabular learning with reinforcement learning for real-world NIDS applications.
We trained 13,440 large language models and found that entropy minimization requires only a single unlabeled data and 10 steps optimization to achieve performance improvements comparable to or even greater than those obtained using thousands of data and carefully designed rewards in rule-based reinforcement learning. This striking result may prompt a rethinking of post-training paradigms for large language models. Our code is avaliable at https://github.com/zitian-gao/one-shot-em.