Multi-agent reinforcement learning (MARL) has achieved significant progress in solving complex multi-player games through self-play. However, training effective adversarial policies requires millions of experience samples and substantial computational resources. Moreover, these policies lack interpretability, hindering their practical deployment. Recently, researchers have successfully leveraged Large Language Models (LLMs) to generate programmatic policies for single-agent tasks, transforming neural network-based policies into interpretable rule-based code with high execution efficiency. Inspired by this, we propose PolicyEvolve, a general framework for generating programmatic policies in multi-player games. PolicyEvolve significantly reduces reliance on manually crafted policy code, achieving high-performance policies with minimal environmental interactions. The framework comprises four modules: Global Pool, Local Pool, Policy Planner, and Trajectory Critic. The Global Pool preserves elite policies accumulated during iterative training. The Local Pool stores temporary policies for the current iteration; only sufficiently high-performing policies from this pool are promoted to the Global Pool. The Policy Planner serves as the core policy generation module. It samples the top three policies from the Global Pool, generates an initial policy for the current iteration based on environmental information, and refines this policy using feedback from the Trajectory Critic. Refined policies are then deposited into the Local Pool. This iterative process continues until the policy achieves a sufficiently high average win rate against the Global Pool, at which point it is integrated into the Global Pool. The Trajectory Critic analyzes interaction data from the current policy, identifies vulnerabilities, and proposes directional improvements to guide the Policy Planner

The rapid expansion of scientific literature makes it increasingly difficult to acquire new knowledge, particularly in specialized domains where reasoning is complex, full-text access is restricted, and target references are sparse among a large set of candidates. We present a Deep Reinforcement Learning framework for sparse reference selection that emulates human knowledge construction, prioritizing which papers to read under limited time and cost. Evaluated on drug--gene relation discovery with access restricted to titles and abstracts, our approach demonstrates that both humans and machines can construct knowledge effectively from partial information.

Game-based interactive driving simulations have emerged as versatile platforms for advancing decision-making algorithms in road transport mobility. While these environments offer safe, scalable, and engaging settings for testing driving strategies, ensuring both realism and robust performance amid dynamic and diverse scenarios remains a significant challenge. Recently, the integration of game-based techniques with advanced learning frameworks has enabled the development of adaptive decision-making models that effectively manage the complexities inherent in varied driving conditions. These models outperform traditional simulation methods, especially when addressing scenario-specific challenges, ranging from obstacle avoidance on highways and precise maneuvering during on-ramp merging to navigation in roundabouts, unsignalized intersections, and even the high-speed demands of autonomous racing. Despite numerous innovations in game-based interactive driving, a systematic review comparing these approaches across different scenarios is still missing. This survey provides a comprehensive evaluation of game-based interactive driving methods by summarizing recent advancements and inherent roadway features in each scenario. Furthermore, the reviewed algorithms are critically assessed based on their adaptation of the standard game model and an analysis of their specific mechanisms to understand their impact on decision-making performance. Finally, the survey discusses the limitations of current approaches and outlines promising directions for future research.

Data collection is crucial for learning robust world models in model-based reinforcement learning. The most prevalent strategies are to actively collect trajectories by interacting with the environment during online training or training on offline datasets. At first glance, the nature of learning task-agnostic environment dynamics makes world models a good candidate for effective offline training. However, the effects of online vs. offline data on world models and thus on the resulting task performance have not been thoroughly studied in the literature. In this work, we investigate both paradigms in model-based settings, conducting experiments on 31 different environments. First, we showcase that online agents outperform their offline counterparts. We identify a key challenge behind performance degradation of offline agents: encountering Out-Of-Distribution states at test time. This issue arises because, without the self-correction mechanism in online agents, offline datasets with limited state space coverage induce a mismatch between the agent's imagination and real rollouts, compromising policy training. We demonstrate that this issue can be mitigated by allowing for additional online interactions in a fixed or adaptive schedule, restoring the performance of online training with limited interaction data. We also showcase that incorporating exploration data helps mitigate the performance degradation of offline agents. Based on our insights, we recommend adding exploration data when collecting large datasets, as current efforts predominantly focus on expert data alone.

We present a Reinforcement Learning (RL)-based locomotion system for Cosmo, a custom-built humanoid robot designed for entertainment applications. Unlike traditional humanoids, entertainment robots present unique challenges due to aesthetic-driven design choices. Cosmo embodies these with a disproportionately large head (16% of total mass), limited sensing, and protective shells that considerably restrict movement. To address these challenges, we apply Adversarial Motion Priors (AMP) to enable the robot to learn natural-looking movements while maintaining physical stability. We develop tailored domain randomization techniques and specialized reward structures to ensure safe sim-to-real, protecting valuable hardware components during deployment. Our experiments demonstrate that AMP generates stable standing and walking behaviors despite Cosmo's extreme mass distribution and movement constraints. These results establish a promising direction for robots that balance aesthetic appeal with functional performance, suggesting that learning-based methods can effectively adapt to aesthetic-driven design constraints.

Solving long-horizon goal-conditioned tasks remains a significant challenge in reinforcement learning (RL). Hierarchical reinforcement learning (HRL) addresses this by decomposing tasks into more manageable sub-tasks, but the automatic discovery of the hierarchy and the joint training of multi-level policies often suffer from instability and can lack theoretical guarantees. In this paper, we introduce Reinforcement Learning with Anticipation (RLA), a principled and potentially scalable framework designed to address these limitations. The RLA agent learns two synergistic models: a low-level, goal-conditioned policy that learns to reach specified subgoals, and a high-level anticipation model that functions as a planner, proposing intermediate subgoals on the optimal path to a final goal. The key feature of RLA is the training of the anticipation model, which is guided by a principle of value geometric consistency, regularized to prevent degenerate solutions. We present proofs that RLA approaches the globally optimal policy under various conditions, establishing a principled and convergent method for hierarchical planning and execution in long-horizon goal-conditioned tasks.

Reinforcement Learning (RL) has shown great potential for autonomous decision-making in the cybersecurity domain, enabling agents to learn through direct environment interaction. However, RL agents in Autonomous Cyber Operations (ACO) typically learn from scratch, requiring them to execute undesirable actions to learn their consequences. In this study, we integrate external knowledge in the form of a Large Language Model (LLM) pretrained on cybersecurity data that our RL agent can directly leverage to make informed decisions. By guiding initial training with an LLM, we improve baseline performance and reduce the need for exploratory actions with obviously negative outcomes. We evaluate our LLM-integrated approach in a simulated cybersecurity environment, and demonstrate that our guided agent achieves over 2x higher rewards during early training and converges to a favorable policy approximately 4,500 episodes faster than the baseline.

Large reasoning models (LRMs) have exhibited strong performance on complex reasoning tasks, with further gains achievable through increased computational budgets at inference. However, current test-time scaling methods predominantly rely on redundant sampling, ignoring the historical experience utilization, thereby limiting computational efficiency. To overcome this limitation, we propose Sticker-TTS, a novel test-time scaling framework that coordinates three collaborative LRMs to iteratively explore and refine solutions guided by historical attempts. At the core of our framework are distilled key conditions-termed stickers-which drive the extraction, refinement, and reuse of critical information across multiple rounds of reasoning. To further enhance the efficiency and performance of our framework, we introduce a two-stage optimization strategy that combines imitation learning with self-improvement, enabling progressive refinement. Extensive evaluations on three challenging mathematical reasoning benchmarks, including AIME-24, AIME-25, and OlymMATH, demonstrate that Sticker-TTS consistently surpasses strong baselines, including self-consistency and advanced reinforcement learning approaches, under comparable inference budgets. These results highlight the effectiveness of sticker-guided historical experience utilization. Our code and data are available at https://github.com/RUCAIBox/Sticker-TTS.

Rechargeable lithium-ion (Li-ion) batteries are a ubiquitous element of modern technology. In the last decades, the production and design of such batteries and their adjacent embedded charging and safety protocols, denoted by Battery Management Systems (BMS), has taken central stage. A fundamental challenge to be addressed is the trade-off between the speed of charging and the ageing behavior, resulting in the loss of capacity in the battery cell. We rely on a high-fidelity physics-based battery model and propose an approach to data-driven charging and safety protocol design. Following a Counterexample-Guided Inductive Synthesis scheme, we combine Reinforcement Learning (RL) with recent developments in data-driven formal methods to obtain a hybrid control strategy: RL is used to synthesise the individual controllers, and a data-driven abstraction guides their partitioning into a switched structure, depending on the initial output measurements of the battery. The resulting discrete selection among RL-based controllers, coupled with the continuous battery dynamics, realises a hybrid system. When a design meets the desired criteria, the abstraction provides probabilistic guarantees on the closed-loop performance of the cell.

Reinforcement Learning from Verifiable Rewards (RLVR) has emerged as a promising framework for enhancing the reasoning capabilities of large language models. However, existing approaches such as GRPO often suffer from zero gradients. This problem arises primarily due to fixed clipping bounds for token-level probability ratios and the standardization of identical rewards, which can lead to ineffective gradient updates and underutilization of generated responses. In this work, we propose Dynamic Clipping Policy Optimization(DCPO), which introduces a dynamic clipping strategy that adaptively adjusts clipping bounds based on token-specific prior probabilities to enhance token-level exploration, and a smooth advantage standardization technique that standardizes rewards across cumulative training steps to improve the response-level effective utilization of generated responses. DCPO achieved state-of-the-art performance on four benchmarks based on four different models. In particular, DCPO achieved an Avg@1 of 46.7 under greedy decoding and an Avg@32 of 38.8 under 32 times sampling on the AIME24 benchmark, surpassing DAPO (36.7/31.6), GRPO (36.7/32.1) and GSPO (40.0/34.9) on the Qwen2.5-Math-7B model. On the AIME25 benchmark based on Qwen2.5-14B, DCPO achieves a performance of (23.3/19.0), surpassing GRPO (13.3/10.5), DAPO (20.0/15.3) and GSPO (16.7/9.9). Furthermore, DCPO achieved an average 28% improvement in the nonzero advantage over GRPO in four models, doubled the training efficiency over DAPO, and significantly reduced the token clipping ratio by an order of magnitude compared to both GRPO and DAPO, while achieving superior performance. These results highlight DCPO's effectiveness in leveraging generated data more efficiently for reinforcement learning in large language models.