Language model self-improvement (LMSI) techniques have recently gained significant attention as they improve language models without requiring external supervision. A notable approach is reinforcement learning from AI feedback (RLAIF), which trains a reward model based on AI preference data and employs reinforcement learning (RL) algorithm to train the language model. However, RLAIF relies on a heuristic assumption that the AI model is able to provide effective feedback, which requires the language model to possess solid capability. In this paper, we present a novel LMSI method, Reinforcement Learning Contemplation (RLC). We disclose that it is simpler for language models to evaluate the text than to generate it, even for small models under 1B parameters. Leveraging the gap between the evaluation and generation, RLC evaluates the generated answers and updates language model using RL to maximize self-evaluation scores. We demonstrate the effectiveness of RLC on a wide range of challenging tasks, including reasoning, summarization, conditioned generation and emotion recognition, resulting in a increase in the answering accuracy (31.23% -> 37.09%) for BigBench-hard reasoning tasks, and a rise in BERTScore for CNN/Daily Mail summarization tasks. In addition, RLC can be applied to models of different sizes (80M to 3B) and model structures (FLAN-T5, LLAMA-3.2 and QWEN-2.5), showcasing its broad applicability. We further verify that when training on larger scale of dataset, RLC improves language model's evaluation and generation ability on unseen tasks simultaneously, enabling a general capability improvement without external supervision.

Hand-craft reward engineering requires domain knowledge with numerous trials and errors, while Preference-based Reinforcement Learning (PbRL) avoids manual reward design but often suffers from limited interpretability and unstable training. To address these issues, we propose a novel preference alignment framework. Our approach leverages large language models to generate sub-reward functions informed by prior knowledge and further align human preferences by optimizing the weights combining these sub-rewards. For policy learning, we introduce Policy Optimization via Pareto Regularization (POPR) which regularizes updates along Pareto-optimal directions. Experiments show that our framework improves reward quality and policy stability, achieving superior performance to expert-designed rewards across most tasks.

Humans possess a large reachable space in the 3D world, enabling interaction with objects at varying heights and distances. However, realizing such large-space reaching on humanoids is a complex whole-body control problem and requires the robot to master diverse skills simultaneously-including base positioning and reorientation, height and body posture adjustments, and end-effector pose control. Learning from scratch often leads to optimization difficulty and poor sim2real transferability. To address this challenge, we propose Real-world-Ready Skill Space (R2S2). Our approach begins with a carefully designed skill library consisting of real-world-ready primitive skills. We ensure optimal performance and robust sim2real transfer through individual skill tuning and sim2real evaluation. These skills are then ensembled into a unified latent space, serving as a structured prior that helps task execution in an efficient and sim2real transferable manner. A high-level planner, trained to sample skills from this space, enables the robot to accomplish real-world goal-reaching tasks. We demonstrate zero-shot sim2real transfer and validate R2S2 in multiple challenging goal-reaching scenarios.

Agentic presentation generation must preserve source content, maintain coherent visual design, render specialized objects, and produce usable artifacts. Existing systems meet only part of this requirement: templates preserve regularity but restrict adaptation, whereas free-form HTML or SVG gives models flexibility at the cost of low-level rendering decisions. This mismatch makes long technical decks brittle, especially when slides contain formulas, code, or data graphics. We present SeaSlides, an agentic slide-generation framework built around a semantic abstraction layer. Rather than authoring coordinates, inline styles, or raw SVG geometry, the model writes structured slide content through reusable components and capability modules, while templates own layout, style, and rendering. We instantiate this principle separately in HTML and Typst: SeaSlides-HTML uses template-defined DOM components, whereas SeaSlides-Typst uses template functions and package-backed modules. Capability modules route equations, code, and charts to dedicated renderers, and three feedback stages localize build errors, project-constraint violations, and visual defects before export. The two systems retain backend-specific syntax and contracts while sharing the same authoring boundary. For evaluation, we combine the 128-task UltraPresent validation setting with SeaSlidesBench-Rich, a new 32-task benchmark stressing mathematics, code, pseudocode, tables, charts, and diagrams. Across four generation models, both SeaSlides backends produce more readable, content-oriented source than SVG-heavy generation. A SeaSlides backend attains the highest rich-content macro-average under three of the four models while maintaining competitive overall qualitative performance. These results support semantic abstraction as a practical authoring principle across presentation backends.

Reinforcement learning (RL) has emerged as a dominant paradigm for eliciting long-horizon reasoning in Large Language Models (LLMs). However, scaling Tool-Integrated Reasoning (TIR) via RL remains challenging due to interaction collapse: a pathological state where models fail to sustain multi-turn tool usage, instead degenerating into heavy internal reasoning with only trivial, post-hoc code verification. We systematically study three questions: (i) how cold-start SFT induces an agentic, tool-using behavioral prior, (ii) how the interaction density of cold-start trajectories shapes exploration and downstream RL outcomes, and (iii) how the RL interaction budget affects learning dynamics and generalization under varying inference-time budgets. We then introduce ASTER (Agentic Scaling with Tool-integrated Extended Reasoning), a framework that circumvents this collapse through a targeted cold-start strategy prioritizing interaction-dense trajectories. We find that a small expert cold-start set of just 4K interaction-dense trajectories yields the strongest downstream performance, establishing a robust prior that enables superior exploration during extended RL training. Extensive evaluations demonstrate that ASTER-4B achieves state-of-the-art results on competitive mathematical benchmarks, reaching 90.0% on AIME 2025, surpassing leading frontier open-source models, including DeepSeek-V3.2-Exp.

Reinforcement learning (RL) has shown promising performance in autonomous driving, yet ensuring the safety of online RL policies remains challenging due to insufficient exposure to safety-critical driving scenes. The long-tailed nature of real-world traffic situations makes dangerous and rare interactions difficult to encounter through conventional sampling, limiting the ability of RL policies to learn robust safety behaviors. Existing methods improve training diversity by synthesizing challenging scenes or adversarial situations. However, these approaches typically optimize scene generation objectives separately from the evolving policy, without explicitly modeling how generated perturbations relate to the current policy's weaknesses and learning needs. In this paper, we propose Threat-guided Policy-aware Scene Perturbation (TPSP) for safe autonomous driving with online RL. TPSP introduces a policy-aware scene encoder to capture the interaction between policy behaviors and surrounding environments, enabling scene perturbation aligned with the current policy. Based on this representation, TPSP selectively perturbs critical objects rather than applying uniform modifications across the scene. Furthermore, we develop a threat-guided optimization strategy that evaluates perturbed scenes through threat-level differences between policy rollouts on original and perturbed scenes, guiding the generation of safety-critical scenes with higher training value. Comprehensive experiments demonstrate that TPSP improves safety learning efficiency, achieving strong safety performance on NAVSIM v2 with approximately 4 million kilometers of simulated driving data. Ablation studies verify that policy-aware targeted perturbations provide more informative safety-critical experiences than random or policy-unaware strategies, enabling safer driving under limited interaction budgets.

Large transformer models, trained on diverse datasets, have demonstrated impressive few-shot performance on previously unseen tasks without requiring parameter updates. This capability has also been explored in Reinforcement Learning (RL), where agents interact with the environment to retrieve context and maximize cumulative rewards, showcasing strong adaptability in complex settings. However, in cooperative Multi-Agent Reinforcement Learning (MARL), where agents must coordinate toward a shared goal, decentralized policy deployment can lead to mismatches in task alignment and reward assignment, limiting the efficiency of policy adaptation. To address this challenge, we introduce Multi-agent In-context Coordination via Decentralized Memory Retrieval (MAICC), a novel approach designed to enhance coordination by fast adaptation. Our method involves training a centralized embedding model to capture fine-grained trajectory representations, followed by decentralized models that approximate the centralized one to obtain team-level task information. Based on the learned embeddings, relevant trajectories are retrieved as context, which, combined with the agents' current sub-trajectories, inform decision-making. During decentralized execution, we introduce a novel memory mechanism that effectively balances test-time online data with offline memory. Based on the constructed memory, we propose a hybrid utility score that incorporates both individual- and team-level returns, ensuring credit assignment across agents. Extensive experiments on cooperative MARL benchmarks, including Level-Based Foraging (LBF) and SMAC (v1/v2), show that MAICC enables faster adaptation to unseen tasks compared to existing methods.

World models enable learning policies via latent imagination, offering benefits such as history compression and sample efficiency. The primary challenge in applying world models to multi-agent tasks is that modeling multi-agent dynamics in latent space requires integrating information from different agents, often creating spurious correlations between their latent states. Existing methods either reconstruct the observation for each agent or employ communication to maintain correlation during execution, failing to learn disentangled latent states that are crucial for effective decentralized control. To address this, we present the Disentangled Multi-Agent World Model (DMAWM). It facilitates learning decentralized policies in the latent space through a novel architecture comprising independent agent modules and a shared environment module. During real-environment execution, agent modules independently process local information to form a factorized latent representation. The environment module is then trained to mirror the factorized structure generated by the agent modules, effectively disentangling individual latent states from the interaction dynamics. Consequently, imaginary rollouts generated by the environment module more faithfully simulate decentralized execution dynamics, facilitating the transfer of policies from imagination to decentralized execution. Empirically, DMAWM outperforms existing model-based and model-free approaches in convergence speed and final performance, with additional visualization demonstrating its efficacy in capturing agent interactions.

Learning reward models from human preference datasets and subsequently optimizing language models via reinforcement learning has emerged as a fundamental paradigm for aligning LLMs with human preferences. The performance of the reward model plays a crucial role in the effectiveness of alignment. Previous reward models operate at a coarse-grained level, requiring the generation of a complete response to obtain a reward value. The sparse reward may present challenges for downstream reinforcement learning. While recent efforts have attempted to learn token-level reward models, the lack of explicit semantic information makes it difficult to model the credit of every individual token. In this paper, we propose assigning scores to every sentence, introducing an intermediate-grained reward model. By segmenting the complete response into sentences and applying differential operations to reward output at the start and end positions of each sentence, we can effectively model the rewards of sentences. Moreover, a novel attention mechanism is introduced to aggregate the scores of all sentences into a response-level score, which allows it to be trained using the Bradley-Terry model. On common benchmarks, our method outperforms the response-level reward model by 2.7 reward modeling evaluation) and surpasses all baselines on AlpacaEval (for alignment evaluation).

Evaluating reward models is a fundamental challenge in Reinforcement Learning (RL), particularly in settings where the reward model is learned or manually designed. The standard paradigm for Reward Model Evaluation (RME) involves training an optimal policy via RL on the given reward model and assessing model quality through the performance of the resulting policy. However, this approach conflates the quality of the reward model with the effectiveness of RL training, and is computationally expensive due to the need for policy optimization. Recent RME methods attempt to circumvent this issue by evaluating reward models directly, without RL, but often rely on impractical assumptions such as access to a ground-truth reward or fail to utilize available supervision in a fine-grained manner. To overcome these limitations, we propose the Policy Preference Alignment Coefficient (PPAC), a novel metric for RME that requires neither RL training nor ground-truth rewards. PPAC first generates a sequence of automatically ranked policy preferences that guarantee monotonic improvement in the policy value, and then quantifies the alignment between these generated preferences and those implied by the candidate reward model. Experimental results across gridworld and continuous control task demonstrate that PPAC yields preference sequences with consistently increasing policy values and outperforms existing metrics in evaluating reward model quality.

Offline Meta-Reinforcement Learning (OMRL) leverages pre-collected data to adapt to new tasks. Context-based methods learn task representations from contexts. However, the context is influenced by both the task and the behavior policy. The mismatch between the behavior policy and the testing policy causes a context distribution shift problem, which results in poor task representations and degraded performance. This problem is exacerbated in settings with data limitations. To address this, we propose a novel approach called Meta-Normalizing Flow (Meta-NF). First, it employs a highly expressive and sample-efficient normalizing flow policy. Second, it incorporates a metric for testing-time task representation selection to effectively mitigate the context shift problem. Empirical results demonstrate that Meta-NF outperforms existing OMRL methods, with both components contributing to its strong performance.

Black-Box Optimization (BBO) has found successful applications in many fields of science and engineering. Recently, there has been a growing interest in meta-learning particular components of BBO algorithms to speed up optimization and get rid of tedious hand-crafted heuristics. As an extension, learning the entire algorithm from data requires the least labor from experts and can provide the most flexibility. In this paper, we propose RIBBO, a method to reinforce-learn a BBO algorithm from offline data in an end-to-end fashion. RIBBO employs expressive sequence models to learn the optimization histories produced by multiple behavior algorithms and tasks, leveraging the in-context learning ability of large models to extract task information and make decisions accordingly. Central to our method is to augment the optimization histories with regret-to-go tokens, which are designed to represent the performance of an algorithm based on cumulative regret over the future part of the histories. The integration of regret-to-go tokens enables RIBBO to automatically generate sequences of query points that are positively correlated to the user-desired regret, verified by its universally good empirical performance on diverse problems, including BBO benchmark, hyper-parameter optimization, and robot control problems.

In reinforcement learning (RL), agents continually interact with the environment and use the feedback to refine their behavior. To guide policy optimization, reward models are introduced as proxies of the desired objectives, such that when the agent maximizes the accumulated reward, it also fulfills the task designer's intentions. Recently, significant attention from both academic and industrial researchers has focused on developing reward models that not only align closely with the true objectives but also facilitate policy optimization. In this survey, we provide a comprehensive review of reward modeling techniques within the deep RL literature. We begin by outlining the background and preliminaries in reward modeling. Next, we present an overview of recent reward modeling approaches, categorizing them based on the source, the mechanism, and the learning paradigm. Building on this understanding, we discuss various applications of these reward modeling techniques and review methods for evaluating reward models. Finally, we conclude by highlighting promising research directions in reward modeling. Altogether, this survey includes both established and emerging methods, filling the vacancy of a systematic review of reward models in current literature.

Utilizing messages from teammates can improve coordination in cooperative multiagent reinforcement learning (MARL). Previous works typically combine raw messages of teammates with local information as inputs for policy. However, neglecting message aggregation poses significant inefficiency for policy learning. Motivated by recent advances in representation learning, we argue that efficient message aggregation is essential for good coordination in cooperative MARL. In this article, we propose M ultiagent communication via S elf-supervised I nformation A ggregation (MASIA), where agents can aggregate the received messages into compact representations with high relevance to augment the local policy. Specifically, we design a permutation-invariant message encoder to generate common information-aggregated representation from messages and optimize it via reconstructing and shooting future information in a self-supervised manner. Hence, each agent would utilize the most relevant parts of the aggregated representation for decision-making by a novel message extraction mechanism. Furthermore, considering the potential of offline learning for real-world applications, we build offline benchmarks for multiagent communication, which is the first as we know. Empirical results demonstrate the superiority of our method in both online and offline settings. We also release the built offline benchmarks in this article as a testbed for communication ability validation to facilitate further future research in this direction.

We introduce EvaLearn, a pioneering benchmark designed to evaluate large language models (LLMs) on their learning capability and efficiency in challenging tasks, a critical, yet underexplored aspect of model potential. EvaLearn contains 648 challenging problems across six task types, grouped into 182 sequences, each sequence dedicated to one task type. Diverging from most existing benchmarks that evaluate models in parallel, EvaLearn requires models to solve problems sequentially, allowing them to leverage the experience gained from previous solutions. EvaLearn provides five comprehensive automated metrics to evaluate models and quantify their learning capability and efficiency. We extensively benchmark nine frontier models and observe varied performance profiles: some models, such as Claude-3.7-sonnet, start with moderate initial performance but exhibit strong learning ability, while some models struggle to benefit from experience and may even show negative transfer. Moreover, we investigate model performance under two learning settings and find that instance-level rubrics and teacher-model feedback further facilitate model learning. Importantly, we observe that current LLMs with stronger static abilities do not show a clear advantage in learning capability across all tasks, highlighting that EvaLearn evaluates a new dimension of model performance. We hope EvaLearn provides a novel evaluation perspective for assessing LLM potential and understanding the gap between models and human capabilities, promoting the development of deeper and more dynamic evaluation approaches. All datasets, the automatic evaluation framework, and the results studied in this paper are available at the GitHub repository.

Low sample efficiency is an enduring challenge of reinforcement learning (RL). With the advent of versatile large language models (LLMs), recent works impart common-sense knowledge to accelerate policy learning for RL processes. However, we note that such guidance is often tailored for one specific task but loses generalizability. In this article, we introduce a framework that harnesses LLMs to extract background knowledge of an environment, which contains general understandings of the entire environment, making various downstream RL tasks benefit from one-time knowledge representation. We ground LLMs by feeding a few precollected experiences and requesting them to delineate background knowledge of the environment. Afterward, we represent the output knowledge as potential functions for potential-based reward shaping, which has a good property for maintaining policy optimality from task rewards. We instantiate three variants to prompt LLMs for background knowledge, including writing code, annotating pReferences, and assigning goals. Our experiments show that these methods achieve significant sample efficiency improvements in a spectrum of downstream tasks from Minigrid and Crafter domains.

Behavior regularization, which constrains the policy to stay close to some behavior policy, is widely used in offline reinforcement learning (RL) to manage the risk of hazardous exploitation of unseen actions. Nevertheless, existing literature on behavior-regularized RL primarily focuses on explicit policy parameterizations, such as Gaussian policies. Consequently, it remains unclear how to extend this framework to more advanced policy parameterizations, such as diffusion models. In this paper, we introduce BDPO, a principled behavior-regularized RL framework tailored for diffusion-based policies, thereby combining the expressive power of diffusion policies and the robustness provided by regularization. The key ingredient of our method is to calculate the Kullback-Leibler (KL) regularization analytically as the accumulated discrepancies in reverse-time transition kernels along the diffusion trajectory. By integrating the regularization, we develop an efficient two-time-scale actor-critic RL algorithm that produces the optimal policy while respecting the behavior constraint. Comprehensive evaluations conducted on synthetic 2D tasks and continuous control tasks from the D4RL benchmark validate its effectiveness and superior performance.

Large Language Models (LLMs), trained on a large amount of corpus, have demonstrated remarkable abilities. However, it may not be sufficient to directly apply open-source LLMs like Llama to certain real-world scenarios, since most of them are trained for \emph{general} purposes. Thus, the demands for customizing publicly available LLMs emerge, but are currently under-studied. In this work, we consider customizing pre-trained LLMs with new human preferences. Specifically, the LLM should not only meet the new preference but also preserve its original capabilities after customization. Drawing inspiration from the observation that human preference can be expressed as a reward model, we propose to cast LLM customization as optimizing the sum of two reward functions, one of which (denoted as $r_1$) was used to pre-train the LLM while the other (denoted as $r_2$) characterizes the new human preference. The obstacle here is that both reward functions are unknown, making the application of modern reinforcement learning methods infeasible. Thanks to the residual Q-learning framework, we can restore the customized LLM with the pre-trained LLM and the \emph{residual Q-function} without the reward function $r_1$. Moreover, we find that for a fixed pre-trained LLM, the reward function $r_2$ can be derived from the residual Q-function, enabling us to directly learn the residual Q-function from the new human preference data upon the Bradley-Terry model. We name our method Q-Adapter as it introduces an adapter module to approximate the residual Q-function for customizing the pre-trained LLM towards the new preference. Experiments based on the Llama-3.1 model on the DSP dataset and HH-RLHF dataset illustrate the superior effectiveness of Q-Adapter on both retaining existing knowledge and learning new preferences. Our code is available at \url{https://github.com/LAMDA-RL/Q-Adapter}.