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% to 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.

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a powerful approach to enhancing the reasoning capabilities of Large Language Models (LLMs), yet its underlying mechanisms remain insufficiently understood. In this work, we undertake a pioneering exploration of RLVR through the novel perspective of token entropy patterns, comprehensively analyzing how different tokens influence reasoning performance. By examining token entropy patterns in Chain-of-Thought (CoT) reasoning, we observe that only a small fraction (approximately 20\%) of tokens exhibit high entropy, and these tokens semantically act as critical forks that steer the model toward diverse reasoning pathways. We further demonstrate that moderately increasing the entropy of these high-entropy tokens via decoding temperature adjustments leads to improved performance, quantitatively confirming their role as decision points in reasoning. We ultimately refine RLVR by restricting policy gradient updates to these forking tokens. Despite utilizing only 20\% of tokens, our approach achieves comparable performance to full-gradient updates on the Qwen3-8B base model. Moreover, it demonstrates remarkable improvements on the larger Qwen3-32B base model, boosting AIME'25 scores by 11.04 and AIME'24 scores by 7.71. In contrast, training exclusively on the 80\% lowest-entropy tokens leads to a marked decline in performance. These findings indicate that the efficacy of RLVR primarily arises from optimizing the high-entropy tokens that dictate key reasoning directions. Collectively, our results suggest promising avenues for optimizing RLVR algorithms by strategically leveraging the potential of these high-entropy minority tokens to further enhance the reasoning abilities of LLMs.

Human beings can make adaptive decisions in a preparatory manner, i.e., by making preparations in advance, which offers significant advantages in scenarios where both online and offline experiences are expensive and limited. Meanwhile, current reinforcement learning methods commonly rely on numerous environment interactions but hardly obtain generalizable policies. In this paper, we introduce the idea of \textit{rehearsal} into policy optimization, where the agent plans for all possible outcomes in mind and acts adaptively according to actual responses from the environment. To effectively rehearse, we propose ReDM, an algorithm that generates a diverse and eligible set of dynamics models and then rehearse the policy via adaptive training on the generated model set. Rehearsal enables the policy to make decision plans for various hypothetical dynamics and to naturally generalize to previously unseen environments. Our experimental results demonstrate that ReDM is capable of learning a valid policy solely through rehearsal, even with \emph{zero} interaction data. We further extend ReDM to scenarios where limited or mismatched interaction data is available, and our experimental results reveal that ReDM produces high-performing policies compared to other offline RL baselines.

Language model self-improvement (LMSI) techniques have recently gained significant attention as they improve language models without requiring external supervision. A common approach is reinforcement learning from AI feedback (RLAIF), which trains a reward model based on AI preference data and employs a reinforcement learning algorithm to train the language model. However, RLAIF relies on the heuristic assumption that an AI model can provide effective feedback and correct wrong answers, requiring a solid capability of the language model. This paper presents a novel LMSI method, Reinforcement Learning Contemplation (RLC). We disclose that it is simpler for language models to evaluate a sentence than to generate it, even for small language models. Leveraging the gap between the evaluation and generation, RLC evaluates generated answers and updates language model parameters using reinforcement learning to maximize evaluation scores. Through testing on various challenging reasoning tasks and text summarization task, our experiments show that RLC effectively improves language model performance without external supervision, resulting in an answering accuracy increase (from 31.23% to 37.09%) for BigBench-hard reasoning tasks, and a rise in BERTScore for CNN/Daily Mail summarization tasks. Furthermore, RLC can be applied to models of different sizes, showcasing its broad applicability.

In reinforcement learning, it is crucial to have an accurate environment dynamics model to evaluate different policies' value in downstream tasks like offline policy optimization and policy evaluation. However, the learned model is known to be inaccurate in predictions when evaluating target policies different from data-collection policies. In this work, we found that utilizing policy representation for model learning, called policy-conditioned model (PCM) learning, is useful to mitigate the problem, especially when the offline dataset is collected from diversified behavior policies. The reason beyond that is in this case, PCM becomes a meta-dynamics model that is trained to be aware of and focus on the evaluation policies that on-the-fly adjust the model to be suitable to the evaluation policies’ state-action distribution, thus improving the prediction accuracy. Based on that intuition, we propose an easy-to-implement yet effective algorithm of PCM for accurate model learning. We also give a theoretical analysis and experimental evidence to demonstrate the feasibility of reducing value gaps by adapting the dynamics model under different policies. Experiment results show that PCM outperforms the existing SOTA off-policy evaluation methods in the DOPE benchmark by a large margin, and derives significantly better policies in offline policy selection and model predictive control compared with the standard model learning method.

One-shot imitation learning (OSIL) is to learn an imitator agent that can execute multiple tasks with only a single demonstration. In real-world scenario, the environment is dynamic, e.g., unexpected changes can occur after demonstration. Thus, achieving generalization of the imitator agent is crucial as agents would inevitably face situations unseen in the provided demonstrations. While traditional OSIL methods excel in relatively stationary settings, their adaptability to such unforeseen changes, which asking for a higher level of generalization ability for the imitator agents, is limited and rarely discussed. In this work, we present a new algorithm called Deep Demonstration Tracing (DDT). In DDT, we propose a demonstration transformer architecture to encourage agents to adaptively trace suitable states in demonstrations. Besides, it integrates OSIL into a meta-reinforcement-learning training paradigm, providing regularization for policies in unexpected situations. We evaluate DDT on a new navigation task suite and robotics tasks, demonstrating its superior performance over existing OSIL methods across all evaluated tasks in dynamic environments with unforeseen changes. The project page is in https://osil-ddt.github.io.

When humans need to learn a new skill, we can acquire knowledge through written books, including textbooks, tutorials, etc. However, current research for decision-making, like reinforcement learning (RL), has primarily required numerous real interactions with the target environment to learn a skill, while failing to utilize the existing knowledge already summarized in the text. The success of Large Language Models (LLMs) sheds light on utilizing such knowledge behind the books. In this paper, we discuss a new policy learning problem called Policy Learning from tutorial Books (PLfB) upon the shoulders of LLMs’ systems, which aims to leverage rich resources such as tutorial books to derive a policy network. Inspired by how humans learn from books, we solve the problem via a three-stage framework: Understanding, Rehearsing, and Introspecting (URI). In particular, it first rehearses decision-making trajectories based on the derived knowledge after understanding the books, then introspects in the imaginary dataset to distill a policy network. We build two benchmarks for PLfB~based on Tic-Tac-Toe and Football games. In experiment, URI's policy achieves at least 44% net win rate against GPT-based agents without any real data; In Football game, which is a complex scenario, URI's policy beat the built-in AIs with a 37% while using GPT-based agent can only achieve a 6\% winning rate. The project page: https://plfb-football.github.io.

Reinforcement learning (RL) traditionally trains agents using interaction data, which limits their capabilities to the scope of the training data. To create more knowledgeable agents, leveraging knowledge from large language models (LLMs) has shown a promising way. Despite various attempts to combine LLMs with RL, there is commonly a semantic gap between action signals and LLM tokens, which hinders their integration. This paper introduces a novel approach, KALM (Knowledgeable Agents from Language Model Rollouts), to learn knowledgeable agents by bridging this gap. KALM extracts knowledge from LLMs in the form of imaginary rollouts, which agents can learn through offline RL. To overcome the limitation that LLMs are inherently text-based and may be incompatible with numerical environmental data, KALM fine-tunes the LLM to perform bidirectional translation between textual goals and rollouts. This process enables the LLM to understand the environment better, facilitating the generation of meaningful rollouts. Experiments on robotic manipulation tasks demonstrate that KALM allows agents to rephrase complex goals and tackle novel tasks requiring new optimal behaviors. KALM achieves a 46% success rate in completing 1400 various novel goals, significantly outperforming the 26% success rate of baseline methods. Project homepage: https://kalmneurips2024.github.io.

Long-term user engagement (LTE) optimization in sequential recommender systems (SRS) is shown to be suited by reinforcement learning (RL) which finds a policy to maximize long-term rewards. Meanwhile, RL has its shortcomings, particularly requiring a large number of online samples for exploration, which is risky in real-world applications. One of the appealing ways to avoid the risk is to build a simulator and learn the optimal recommendation policy in the simulator. In LTE optimization, the simulator is to simulate multiple users’ daily feedback for given recommendations. However, building a user simulator with no reality-gap, i.e., can predict user’s feedback exactly, is unrealistic because the users’ reaction patterns are complex and historical logs for each user are limited, which might mislead the simulator-based recommendation policy. In this paper, we present a practical simulator-based recommender policy training approach, Simulation-to-Recommendation (Sim2Rec) to handle the reality-gap problem for LTE optimization. Specifically, Sim2Rec introduces a simulator set to generate various possibilities of user behavior patterns, then trains an environment-parameter extractor to recognize users’ behavior patterns in the simulators. Finally, a context-aware policy is trained to make the optimal decisions on all of the variants of the users based on the inferred environment-parameters. The policy is transferable to unseen environments (e.g., the real world) directly as it has learned to recognize all various user behavior patterns and to make the correct decisions based on the inferred environment-parameters. Experiments are conducted in synthetic environments and a real-world large-scale ride-hailing platform, DidiChuxing. The results show that Sim2Rec achieves significant performance improvement, and produces robust recommendations in unseen environments.

Domestic service robots are becoming increasingly popular due to their ability to help people with household tasks. These robots often encounter the challenge of manipulating objects in cluttered environments (MoC), which is difficult due to the complexity of effective planning and control. Previous solutions involved designing specific action primitives and planning paradigms. However, the pre-coded action primitives can limit the agility and task-solving scope of robots. In this paper, we propose a general approach for MoC called the Object-Oriented Option Framework (O3F), which uses the option framework (OF) to learn planning and control. The standard OF discovers options from scratch based on reinforcement learning, which can lead to collapsed options and hurt learning. To address this limitation, O3F introduces the concept of an object-oriented option space for OF, which focuses specifically on object movement and overcomes the challenges associated with collapsed options. Based on this, we train an object-oriented option planner to determine the option to execute and a universal object-oriented option executor to complete the option. Simulation experiments on the Ginger XR1 robot and robot arm show that O3F is generally applicable to various types of robot and manipulation tasks. Furthermore, O3F achieves success rates of 72.4% and 90% in grasping and object collecting tasks, respectively, significantly outperforming baseline methods.

An accurate environment dynamics model is crucial for various downstream tasks in sequential decision-making, such as counterfactual prediction, off-policy evaluation, and offline reinforcement learning. Currently, these models were learned through empirical risk minimization (ERM) by step-wise fitting of historical transition data. This way was previously believed unreliable over long-horizon rollouts because of the compounding errors, which can lead to uncontrollable inaccuracies in predictions. In this paper, we find that the challenge extends beyond just long-term prediction errors: we reveal that even when planning with one step, learned dynamics models can also perform poorly due to the selection bias of behavior policies during data collection. This issue will significantly mislead the policy optimization process even in identifying single-step optimal actions, further leading to a greater risk in sequential decision-making scenarios. To tackle this problem, we introduce a novel model-learning objective called adversarial weighted empirical risk minimization (AWRM). AWRM incorporates an adversarial policy that exploits the model to generate a data distribution that weakens the model's prediction accuracy, and subsequently, the model is learned under this adversarial data distribution. We implement a practical algorithm, GALILEO, for AWRM and evaluate it on two synthetic tasks, three continuous-control tasks, and \textit{a real-world application}. The experiments demonstrate that GALILEO can accurately predict counterfactual actions and improve various downstream tasks, including offline policy evaluation and improvement, as well as online decision-making.

Natural language-conditioned reinforcement learning (RL) enables agents to follow human instructions. Previous approaches generally implemented language-conditioned RL by providing the policy with human instructions in natural language (NL) and training the policy to follow instructions. In this is outside-in approach, the policy must comprehend the NL and manage the task simultaneously. However, the unbounded NL examples often bring much extra complexity for solving concrete RL tasks, which can distract policy learning from completing the task. To ease the learning burden of the policy, we investigate an inside-out scheme for natural language-conditioned RL by developing a task language (TL) that is task-related and easily understood by the policy, thus reducing the policy learning burden. Besides, we employ a translator to translate natural language into the TL, which is used in RL to achieve efficient policy training. We implement this scheme as TALAR (TAsk Language with predicAte Representation) that learns multiple predicates to model object relationships as the TL. Experiments indicate that TALAR not only better comprehends NL instructions but also leads to a better instruction-following policy that significantly improves the success rate over baselines and adapts to unseen expressions of NL instruction. Besides, the TL is also an effective sub-task abstraction compatible with hierarchical RL.

In reinforcement learning, a promising direction to avoid online trial-and-error costs is learning from an offline dataset. Current offline reinforcement learning methods commonly learn in the policy space constrained to in-support regions by the offline dataset, in order to ensure the robustness of the outcome policies. Such constraints, however, also limit the potential of the outcome policies. In this paper, to release the potential of offline policy learning, we investigate the decision-making problems in out-of-support regions directly and propose offline Model-based Adaptable Policy LEarning (MAPLE). By this approach, instead of learning in in-support regions, we learn an adaptable policy that can adapt its behavior in out-of-support regions when deployed. We give a practical implementation of MAPLE via meta-learning techniques and ensemble model learning techniques. We conduct experiments on MuJoCo locomotion tasks with offline datasets. The results show that the proposed method can make robust decisions in out-of-support regions and achieve better performance than SOTA algorithms.

Offline reinforcement learning (RL) aims at learning effective policies from historical data without extra environment interactions. During our experience of applying offline RL, we noticed that previous offline RL benchmarks commonly involve significant reality gaps, which we have identified include rich and overly exploratory datasets, degraded baseline, and missing policy validation. In many real-world situations, to ensure system safety, running an overly exploratory policy to collect various data is prohibited, thus only a narrow data distribution is available. The resulting policy is regarded as effective if it is better than the working behavior policy; the policy model can be deployed only if it has been well validated, rather than accomplished the training. In this paper, we present a Near real-world offline RL benchmark, named NeoRL, to reflect these properties. NeoRL datasets are collected with a more conservative strategy. Moreover, NeoRL contains the offline training and offline validation pipeline before the online test, corresponding to real-world situations. We then evaluate recent state-of-the-art offline RL algorithms in NeoRL. The empirical results demonstrate that some offline RL algorithms are less competitive to the behavior cloning and the deterministic behavior policy, implying that they could be less effective in real-world tasks than in the previous benchmarks. We also disclose that current offline policy evaluation methods could hardly select the best policy. We hope this work will shed some light on future research and deploying RL in real-world systems.

In visual-input sim-to-real scenarios, to overcome the reality gap between images rendered in simulators and those from the real world, domain adaptation, i.e., learning an aligned representation space between simulators and the real world, then training and deploying policies in the aligned representation, is a promising direction. Previous methods focus on same-modal domain adaptation. However, those methods require building and running simulators that render high-quality images, which can be difficult and costly. In this paper, we consider a more cost-efficient setting of visual-input sim-to-real where only low-dimensional states are simulated. We first point out that the objective of learning mapping functions in previous methods that align the representation spaces is ill-posed, prone to yield an incorrect mapping. When the mapping crosses modalities, previous methods are easier to fail. Our algorithm, Cross-mOdal Domain Adaptation with Sequential structure (CODAS), mitigates the ill-posedness by utilizing the sequential nature of the data sampling process in RL tasks. Experiments on MuJoCo and Hand Manipulation Suite tasks show that the agents deployed with our method achieve similar performance as it has in the source domain, while those deployed with previous methods designed for same-modal domain adaptation suffer a larger performance gap.

In reinforcement learning, a promising direction to avoid online trial-and-error costs is learning from an offline dataset. Current offline reinforcement learning methods commonly learn in the policy space constrained to in-support regions by the offline dataset, in order to ensure the robustness of the outcome policies. Such constraints, however, also limit the potential of the outcome policies. In this paper, to release the potential of offline policy learning, we investigate the decision-making problems in out-of-support regions directly and propose offline Model-based Adaptable Policy LEarning (MAPLE). By this approach, instead of learning in in-support regions, we learn an adaptable policy that can adapt its behavior in out-of-support regions when deployed. We conduct experiments on MuJoCo controlling tasks with offline datasets. The results show that the proposed method can make robust decisions in out-of-support regions and achieve better performance than SOTA algorithms.