Recent advancements in language models have demonstrated remarkable in-context learning abilities, prompting the exploration of in-context reinforcement learning (ICRL) to extend the promise to decision domains. Due to involving more complex dynamics and temporal correlations, existing ICRL approaches may face challenges in learning from suboptimal trajectories and achieving precise in-context inference. In the paper, we propose \textbf{S}calable \textbf{I}n-\textbf{C}ontext \textbf{Q}-\textbf{L}earning (\textbf{SICQL}), an innovative framework that harnesses dynamic programming and world modeling to steer ICRL toward efficient reward maximization and task generalization, while retaining the scalability and stability of supervised pretraining. We design a prompt-based multi-head transformer architecture that simultaneously predicts optimal policies and in-context value functions using separate heads. We pretrain a generalized world model to capture task-relevant information, enabling the construction of a compact prompt that facilitates fast and precise in-context inference. During training, we perform iterative policy improvement by fitting a state value function to an upper-expectile of the Q-function, and distill the in-context value functions into policy extraction using advantage-weighted regression. Extensive experiments across a range of discrete and continuous environments show consistent performance gains over various types of baselines, especially when learning from suboptimal data. Our code is available at https://github.com/NJU-RL/SICQL

Language is a powerful communicative and cognitive tool. It enables humans to express thoughts, share intentions, and reason about complex phenomena. Despite our fluency in using and understanding language, the question of how it arises and evolves over time remains unsolved. A leading hypothesis in linguistics and anthropology posits that language evolved to meet the ecological and social demands of early human cooperation. Language did not arise in isolation, but through shared survival goals. Inspired by this view, we investigate the emergence of language in multi-agent Foraging Games. These environments are designed to reflect the cognitive and ecological constraints believed to have influenced the evolution of communication. Agents operate in a shared grid world with only partial knowledge about other agents and the environment, and must coordinate to complete games like picking up high-value targets or executing temporally ordered actions. Using end-to-end deep reinforcement learning, agents learn both actions and communication strategies from scratch. We find that agents develop communication protocols with hallmark features of natural language: arbitrariness, interchangeability, displacement, cultural transmission, and compositionality. We quantify each property and analyze how different factors, such as population size, social dynamics, and temporal dependencies, shape specific aspects of the emergent language. Our framework serves as a platform for studying how language can evolve from partial observability, temporal reasoning, and cooperative goals in embodied multi-agent settings. We will release all data, code, and models publicly.

Image captioning is a fundamental task that bridges the visual and linguistic domains, playing a critical role in pre-training Large Vision-Language Models (LVLMs). Current state-of-the-art captioning models are typically trained with Supervised Fine-Tuning (SFT), a paradigm that relies on expensive, non-scalable data annotated by humans or proprietary models. This approach often leads to models that memorize specific ground-truth answers, limiting their generality and ability to generate diverse, creative descriptions. To overcome the limitation of SFT, we propose applying the Reinforcement Learning with Verifiable Rewards (RLVR) paradigm to the open-ended task of image captioning. A primary challenge, however, is designing an objective reward function for the inherently subjective nature of what constitutes a "good" caption. We introduce Captioning Reinforcement Learning (CapRL), a novel training framework that redefines caption quality through its utility: a high-quality caption should enable a non-visual language model to accurately answer questions about the corresponding image. CapRL employs a decoupled two-stage pipeline where an LVLM generates a caption, and the objective reward is derived from the accuracy of a separate, vision-free LLM answering Multiple-Choice Questions based solely on that caption. As the first study to apply RLVR to the subjective image captioning task, we demonstrate that CapRL significantly enhances multiple settings. Pretraining on the CapRL-5M caption dataset annotated by CapRL-3B results in substantial gains across 12 benchmarks. Moreover, within the Prism Framework for caption quality evaluation, CapRL achieves performance comparable to Qwen2.5-VL-72B, while exceeding the baseline by an average margin of 8.4%. Code is available here: https://github.com/InternLM/CapRL.

Decentralized Autonomous Organizations (DAOs) use smart contracts to foster communities working toward common goals. Existing definitions of decentralization, however -- the 'D' in DAO -- fall short of capturing the key properties characteristic of diverse and equitable participation. This work proposes a new framework for measuring DAO decentralization called Voting-Bloc Entropy (VBE, pronounced ''vibe''). VBE is based on the idea that voters with closely aligned interests act as a centralizing force and should be modeled as such. VBE formalizes this notion by measuring the similarity of participants' utility functions across a set of voting rounds. Unlike prior, ad hoc definitions of decentralization, VBE derives from first principles: We introduce a simple (yet powerful) reinforcement learning-based conceptual model for voting, that in turn implies VBE. We first show VBE's utility as a theoretical tool. We prove a number of results about the (de)centralizing effects of vote delegation, proposal bundling, bribery, etc. that are overlooked in previous notions of DAO decentralization. Our results lead to practical suggestions for enhancing DAO decentralization. We also show how VBE can be used empirically by presenting measurement studies and VBE-based governance experiments. We make the tools we developed for these results available to the community in the form of open-source artifacts in order to facilitate future study of DAO decentralization.

Learning a control policy for a multi-phase, long-horizon task, such as basketball maneuvers, remains challenging for reinforcement learning approaches due to the need for seamless policy composition and transitions between skills. A long-horizon task typically consists of distinct subtasks with well-defined goals, separated by transitional subtasks with unclear goals but critical to the success of the entire task. Existing methods like the mixture of experts and skill chaining struggle with tasks where individual policies do not share significant commonly explored states or lack well-defined initial and terminal states between different phases. In this paper, we introduce a novel policy integration framework to enable the composition of drastically different motor skills in multi-phase long-horizon tasks with ill-defined intermediate states. Based on that, we further introduce a high-level soft router to enable seamless and robust transitions between the subtasks. We evaluate our framework on a set of fundamental basketball skills and challenging transitions. Policies trained by our approach can effectively control the simulated character to interact with the ball and accomplish the long-horizon task specified by real-time user commands, without relying on ball trajectory references.

Reward design remains a critical bottleneck in visual reinforcement learning (RL) for robotic manipulation. In simulated environments, rewards are conventionally designed based on the distance to a target position. However, such precise positional information is often unavailable in real-world visual settings due to sensory and perceptual limitations. In this study, we propose a method that implicitly infers spatial distances through keypoints extracted from images. Building on this, we introduce Reward Learning with Anticipation Model (ReLAM), a novel framework that automatically generates dense, structured rewards from action-free video demonstrations. ReLAM first learns an anticipation model that serves as a planner and proposes intermediate keypoint-based subgoals on the optimal path to the final goal, creating a structured learning curriculum directly aligned with the task's geometric objectives. Based on the anticipated subgoals, a continuous reward signal is provided to train a low-level, goal-conditioned policy under the hierarchical reinforcement learning (HRL) framework with provable sub-optimality bound. Extensive experiments on complex, long-horizon manipulation tasks show that ReLAM significantly accelerates learning and achieves superior performance compared to state-of-the-art methods.

Reinforcement learning (RL) has demonstrated impressive ability to learn high-performing policies across diverse domains, including gaming (Mnih et al., 2015; Vinyals et al., 2019; Berner et al., 2019), robotics (Levine et al., 2016; Akkaya et al., 2019), traffic control (El-Tantawy et al., 2013), and aligning large language models with human preferences (Stiennon et al., 2020). However, learning a high-performing policy in RL typically requires extensive data collection, thereby discouraging practical deployment. To alleviate the burden of extensive data collection, recent work has explored leveraging priors, either via a pre-collected dataset (Levine et al., 2020) or a reference policy (Xie et al., 2021; Kalashnikov et al., 2018). In practice, however, the deployment task often differs from that represented by the dataset or reference policy; such task mismatch can substantially diminish the utility of these priors. To leverage priors despite this mismatch, several approaches have been proposed in the context of meta-RL (Wang et al., 2016; Duan et al., 2016; Finn et al., 2017; Rakelly et al., 2019), which aim to leverage prior knowledge for efficient adaptation, either by (i) improving the sample efficiency of standard RL or by (ii) enabling rapid adaptation to new tasks.

Equity in real-world sequential decision problems can be enforced using fairness-aware methods. Therefore, we require algorithms that can make suitable and transparent trade-offs between performance and the desired fairness notions. As the desired performance-fairness trade-off is hard to specify a priori, we propose a framework where multiple trade-offs can be explored. Insights provided by the reinforcement learning algorithm regarding the obtainable performance-fairness trade-offs can then guide stakeholders in selecting the most appropriate policy. To capture fairness, we propose an extended Markov decision process, $f$MDP, that explicitly encodes individuals and groups. Given this $f$MDP, we formalise fairness notions in the context of sequential decision problems and formulate a fairness framework that computes fairness measures over time. We evaluate our framework in two scenarios with distinct fairness requirements: job hiring, where strong teams must be composed while treating applicants equally, and fraud detection, where fraudulent transactions must be detected while ensuring the burden on customers is fairly distributed. We show that our framework learns policies that are more fair across multiple scenarios, with only minor loss in performance reward. Moreover, we observe that group and individual fairness notions do not necessarily imply one another, highlighting the benefit of our framework in settings where both fairness types are desired. Finally, we provide guidelines on how to apply this framework across different problem settings.

This study examines the potential impact of reinforcement learning (RL)-enabled autonomous vehicles (AV) on urban traffic flow in a mixed traffic environment. We focus on a simplified day-to-day route choice problem in a multi-agent setting. We consider a city network where human drivers travel through their chosen routes to reach their destinations in minimum travel time. Then, we convert one-third of the population into AVs, which are RL agents employing Deep Q-learning algorithm. We define a set of optimization targets, or as we call them behaviors, namely selfish, collaborative, competitive, social, altruistic, and malicious. We impose a selected behavior on AVs through their rewards. We run our simulations using our in-house developed RL framework PARCOUR. Our simulations reveal that AVs optimize their travel times by up to 5\%, with varying impacts on human drivers' travel times depending on the AV behavior. In all cases where AVs adopt a self-serving behavior, they achieve shorter travel times than human drivers. Our findings highlight the complexity differences in learning tasks of each target behavior. We demonstrate that the multi-agent RL setting is applicable for collective routing on traffic networks, though their impact on coexisting parties greatly varies with the behaviors adopted.

Most modern approaches to quadruped locomotion focus on using Deep Reinforcement Learning (DRL) to learn policies from scratch, in an end-to-end manner. Such methods often fail to scale, as every new problem or application requires time-consuming and iterative reward definition and tuning. We present Generalist Contact-Conditioned Policy (GeCCo) -- a low-level policy trained with Deep Reinforcement Learning that is capable of tracking arbitrary contact points on a quadruped robot. The strength of our approach is that it provides a general and modular low-level controller that can be reused for a wider range of high-level tasks, without the need to re-train new controllers from scratch. We demonstrate the scalability and robustness of our method by evaluating on a wide range of locomotion and manipulation tasks in a common framework and under a single generalist policy. These include a variety of gaits, traversing complex terrains (eg. stairs and slopes) as well as previously unseen stepping-stones and narrow beams, and interacting with objects (eg. pushing buttons, tracking trajectories). Our framework acquires new behaviors more efficiently, simply by combining a task-specific high-level contact planner and the pre-trained generalist policy. A supplementary video can be found at https://youtu.be/o8Dd44MkG2E.