Deheng Ye

Hanchi Huang, Deheng Ye, Li Shen et al.

We introduce CAMRL, the first curriculum-based asymmetric multi-task learning (AMTL) algorithm for dealing with multiple reinforcement learning (RL) tasks altogether. To mitigate the negative influence of customizing the one-off training order in curriculum-based AMTL, CAMRL switches its training mode between parallel single-task RL and asymmetric multi-task RL (MTRL), according to an indicator regarding the training time, the overall performance, and the performance gap among tasks. To leverage the multi-sourced prior knowledge flexibly and to reduce negative transfer in AMTL, we customize a composite loss with multiple differentiable ranking functions and optimize the loss through alternating optimization and the Frank-Wolfe algorithm. The uncertainty-based automatic adjustment of hyper-parameters is also applied to eliminate the need of laborious hyper-parameter analysis during optimization. By optimizing the composite loss, CAMRL predicts the next training task and continuously revisits the transfer matrix and network weights. We have conducted experiments on a wide range of benchmarks in multi-task RL, covering Gym-minigrid, Meta-world, Atari video games, vision-based PyBullet tasks, and RLBench, to show the improvements of CAMRL over the corresponding single-task RL algorithm and state-of-the-art MTRL algorithms. The code is available at: https://github.com/huanghanchi/CAMRL

Jiate Liu, Yiqin Zhu, Kaiwen Xiao et al.

The goal of program synthesis, or code generation, is to generate executable code based on given descriptions. Recently, there has been an increasing number of studies employing reinforcement learning (RL) to improve the performance of large language models (LLMs) for code. However, current representative works either rely solely on offline frameworks, limiting the exploration of new sample spaces, or fall short in the utilization of unit test signals, not accounting for specific error locations within the code. To address these issues, we propose RLTF, i.e., Reinforcement Learning from Unit Test Feedback, a novel online RL framework with unit test feedback of multi-granularity for refining code LLMs. Our approach generates data in real-time during training and simultaneously utilizes fine-grained feedback signals to guide the model towards producing higher-quality code. Extensive experiments show that RLTF achieves state-of-the-art performance on the APPS and the MBPP benchmarks. Our code is available at: https://github.com/Zyq-scut/RLTF.

Yihuai Lan, Zhiqiang Hu, Lei Wang et al.

This paper explores the open research problem of understanding the social behaviors of LLM-based agents. Using Avalon as a testbed, we employ system prompts to guide LLM agents in gameplay. While previous studies have touched on gameplay with LLM agents, research on their social behaviors is lacking. We propose a novel framework, tailored for Avalon, features a multi-agent system facilitating efficient communication and interaction. We evaluate its performance based on game success and analyze LLM agents' social behaviors. Results affirm the framework's effectiveness in creating adaptive agents and suggest LLM-based agents' potential in navigating dynamic social interactions. By examining collaboration and confrontation behaviors, we offer insights into this field's research and applications. Our code is publicly available at https://github.com/3DAgentWorld/LLM-Game-Agent.

Haibin Zhou, Tong Wei, Zichuan Lin et al.

We study the adaption of Soft Actor-Critic (SAC), which is considered as a state-of-the-art reinforcement learning (RL) algorithm, from continuous action space to discrete action space. We revisit vanilla discrete SAC and provide an in-depth understanding of its Q value underestimation and performance instability issues when applied to discrete settings. We thereby propose Stable Discrete SAC (SDSAC), an algorithm that leverages entropy-penalty and double average Q-learning with Q-clip to address these issues. Extensive experiments on typical benchmarks with discrete action space, including Atari games and a large-scale MOBA game, show the efficacy of our proposed method. Our code is at: https://github.com/coldsummerday/SD-SAC.git.

Boxuan Zhao, Jun Zhang, Deheng Ye et al.

Whole-slide images (WSI) in computational pathology have high resolution with gigapixel size, but are generally with sparse regions of interest, which leads to weak diagnostic relevance and data inefficiency for each area in the slide. Most of the existing methods rely on a multiple instance learning framework that requires densely sampling local patches at high magnification. The limitation is evident in the application stage as the heavy computation for extracting patch-level features is inevitable. In this paper, we develop RLogist, a benchmarking deep reinforcement learning (DRL) method for fast observation strategy on WSIs. Imitating the diagnostic logic of human pathologists, our RL agent learns how to find regions of observation value and obtain representative features across multiple resolution levels, without having to analyze each part of the WSI at the high magnification. We benchmark our method on two whole-slide level classification tasks, including detection of metastases in WSIs of lymph node sections, and subtyping of lung cancer. Experimental results demonstrate that RLogist achieves competitive classification performance compared to typical multiple instance learning algorithms, while having a significantly short observation path. In addition, the observation path given by RLogist provides good decision-making interpretability, and its ability of reading path navigation can potentially be used by pathologists for educational/assistive purposes. Our code is available at: \url{https://github.com/tencent-ailab/RLogist}.

Hanchi Huang, Li Shen, Deheng Ye et al.

We propose a novel master-slave architecture to solve the top-$K$ combinatorial multi-armed bandits problem with non-linear bandit feedback and diversity constraints, which, to the best of our knowledge, is the first combinatorial bandits setting considering diversity constraints under bandit feedback. Specifically, to efficiently explore the combinatorial and constrained action space, we introduce six slave models with distinguished merits to generate diversified samples well balancing rewards and constraints as well as efficiency. Moreover, we propose teacher learning based optimization and the policy co-training technique to boost the performance of the multiple slave models. The master model then collects the elite samples provided by the slave models and selects the best sample estimated by a neural contextual UCB-based network to make a decision with a trade-off between exploration and exploitation. Thanks to the elaborate design of slave models, the co-training mechanism among slave models, and the novel interactions between the master and slave models, our approach significantly surpasses existing state-of-the-art algorithms in both synthetic and real datasets for recommendation tasks. The code is available at: \url{https://github.com/huanghanchi/Master-slave-Algorithm-for-Top-K-Bandits}.

Ruize Zhang, Zelai Xu, Chengdong Ma et al. · tsinghua

Self-play, a learning paradigm where agents iteratively refine their policies by interacting with historical or concurrent versions of themselves or other evolving agents, has shown remarkable success in solving complex non-cooperative multi-agent tasks. Despite its growing prominence in multi-agent reinforcement learning (MARL), such as Go, poker, and video games, a comprehensive and structured understanding of self-play remains lacking. This survey fills this gap by offering a comprehensive roadmap to the diverse landscape of self-play methods. We begin by introducing the necessary preliminaries, including the MARL framework and basic game theory concepts. Then, it provides a unified framework and classifies existing self-play algorithms within this framework. Moreover, the paper bridges the gap between the algorithms and their practical implications by illustrating the role of self-play in different non-cooperative scenarios. Finally, the survey highlights open challenges and future research directions in self-play.

Zichuan Lin, Xiapeng Wu, Mingfei Sun et al.

Recent success in Deep Reinforcement Learning (DRL) methods has shown that policy optimization with respect to an off-policy distribution via importance sampling is effective for sample reuse. In this paper, we show that the use of importance sampling could introduce high variance in the objective estimate. Specifically, we show in a principled way that the variance of importance sampling estimate grows quadratically with importance ratios and the large ratios could consequently jeopardize the effectiveness of surrogate objective optimization. We then propose a technique called sample dropout to bound the estimation variance by dropping out samples when their ratio deviation is too high. We instantiate this sample dropout technique on representative policy optimization algorithms, including TRPO, PPO, and ESPO, and demonstrate that it consistently boosts the performance of those DRL algorithms on both continuous and discrete action controls, including MuJoCo, DMControl and Atari video games. Our code is open-sourced at \url{https://github.com/LinZichuan/sdpo.git}.

Yun Qu, Boyuan Wang, Jianzhun Shao et al. · tsinghua

The advancement of Offline Reinforcement Learning (RL) and Offline Multi-Agent Reinforcement Learning (MARL) critically depends on the availability of high-quality, pre-collected offline datasets that represent real-world complexities and practical applications. However, existing datasets often fall short in their simplicity and lack of realism. To address this gap, we propose Hokoff, a comprehensive set of pre-collected datasets that covers both offline RL and offline MARL, accompanied by a robust framework, to facilitate further research. This data is derived from Honor of Kings, a recognized Multiplayer Online Battle Arena (MOBA) game known for its intricate nature, closely resembling real-life situations. Utilizing this framework, we benchmark a variety of offline RL and offline MARL algorithms. We also introduce a novel baseline algorithm tailored for the inherent hierarchical action space of the game. We reveal the incompetency of current offline RL approaches in handling task complexity, generalization and multi-task learning.

Zhifei Xie, Zihang Liu, Ze An et al.

Audio is an inherently interactive modality, yet today's Large Audio Language Models (LALMs) are offline, and streaming audio models each handle only a single task such as streaming ASR or voice chatting. It is time to unify them into one online LALM: a model that, through an always-on perceive-decide-respond loop, listens to sound, environment, and instructions in real time and reacts on the fly. We formalize this regime as the Audio Interaction Model, and realize it with Audio-Interaction, a unified streaming model that retains offline task execution while adding online general audio instruction following, from dialogue to full voice chatting, deciding when to respond from the semantics of the stream. To enable this, we propose SoundFlow, a framework that instantiates the perceive-decide-respond loop end to end, from data to training to deployment, through streaming-native data construction, comprehension-aware training, and asynchronous low-latency inference for stable real-time interaction. We further construct StreamAudio-2M, a 2.6M-item streaming corpus spanning 7 fundamental abilities and 28 sub-tasks, and Proactive-Sound-Bench for evaluating proactive audio intervention. Across 8 benchmarks, Audio-Interaction preserves competitive performance on mainstream audio tasks while unlocking capabilities inaccessible to offline LALMs, including real-time ASR, streaming audio instruction following, and proactive help.

Tao Jiang, Zichuan Lin, Lihe Li et al.

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. Code is available at https://github.com/LAMDA-RL/MAICC.

Tiantian Zhang, Zichuan Lin, Yuxing Wang et al.

A key challenge of continual reinforcement learning (CRL) in dynamic environments is to promptly adapt the RL agent's behavior as the environment changes over its lifetime, while minimizing the catastrophic forgetting of the learned information. To address this challenge, in this article, we propose DaCoRL, i.e., dynamics-adaptive continual RL. DaCoRL learns a context-conditioned policy using progressive contextualization, which incrementally clusters a stream of stationary tasks in the dynamic environment into a series of contexts and opts for an expandable multihead neural network to approximate the policy. Specifically, we define a set of tasks with similar dynamics as an environmental context and formalize context inference as a procedure of online Bayesian infinite Gaussian mixture clustering on environment features, resorting to online Bayesian inference to infer the posterior distribution over contexts. Under the assumption of a Chinese restaurant process prior, this technique can accurately classify the current task as a previously seen context or instantiate a new context as needed without relying on any external indicator to signal environmental changes in advance. Furthermore, we employ an expandable multihead neural network whose output layer is synchronously expanded with the newly instantiated context, and a knowledge distillation regularization term for retaining the performance on learned tasks. As a general framework that can be coupled with various deep RL algorithms, DaCoRL features consistent superiority over existing methods in terms of the stability, overall performance and generalization ability, as verified by extensive experiments on several robot navigation and MuJoCo locomotion tasks.

Zhifei Xie, Kaiyu Pang, Haobin Zhang et al.

Despite rapid advances in automatic speech recognition (ASR) and large audio-language models, robust recognition in real-world environments remains limited by an "acoustic robustness bottleneck": models often lose acoustic grounding and produce omissions or hallucinations under severe, compositional distortions. We propose Mega-ASR, a unified ASR-in-the-wild framework that combines scalable compound-data construction with progressive acoustic-to-semantic optimization. We introduce Voices-in-the-Wild-2M, covering 7 classic acoustic phenomena and 54 physically plausible compound scenarios, and train Mega-ASR with Acoustic-to-Semantic Progressive Supervised Fine-Tuning and Dual-Granularity WER-Gated Policy Optimization. Extensive experiments demonstrate that Mega-ASR achieves significant advantages over prior state-of-the-art systems on adverse-condition ASR benchmarks (45.69% vs. 54.01% on VOiCES R4-B-F, and 21.49% vs. 29.34% on NOIZEUS Sta-0). On complex compositional acoustic scenarios, Mega-ASR further delivers over 30% relative WER reduction against strong open- and closed-source baselines, establishing a scalable paradigm for robust ASR in-the-wild.

Zhihui Xie, Zichuan Lin, Junyou Li et al.

The past few years have seen rapid progress in combining reinforcement learning (RL) with deep learning. Various breakthroughs ranging from games to robotics have spurred the interest in designing sophisticated RL algorithms and systems. However, the prevailing workflow in RL is to learn tabula rasa, which may incur computational inefficiency. This precludes continuous deployment of RL algorithms and potentially excludes researchers without large-scale computing resources. In many other areas of machine learning, the pretraining paradigm has shown to be effective in acquiring transferable knowledge, which can be utilized for a variety of downstream tasks. Recently, we saw a surge of interest in Pretraining for Deep RL with promising results. However, much of the research has been based on different experimental settings. Due to the nature of RL, pretraining in this field is faced with unique challenges and hence requires new design principles. In this survey, we seek to systematically review existing works in pretraining for deep reinforcement learning, provide a taxonomy of these methods, discuss each sub-field, and bring attention to open problems and future directions.

Ke Xu, Jianqiao Wangni, Yifan Zhang et al.

Data explosion and an increase in model size drive the remarkable advances in large-scale machine learning, but also make model training time-consuming and model storage difficult. To address the above issues in the distributed model training setting which has high computation efficiency and less device limitation, there are still two main difficulties. On one hand, the communication costs for exchanging information, e.g., stochastic gradients among different workers, is a key bottleneck for distributed training efficiency. On the other hand, less parameter model is easy for storage and communication, but the risk of damaging the model performance. To balance the communication costs, model capacity and model performance simultaneously, we propose quantized composite mirror descent adaptive subgradient (QCMD adagrad) and quantized regularized dual average adaptive subgradient (QRDA adagrad) for distributed training. To be specific, we explore the combination of gradient quantization and sparse model to reduce the communication cost per iteration in distributed training. A quantized gradient-based adaptive learning rate matrix is constructed to achieve a balance between communication costs, accuracy, and model sparsity. Moreover, we theoretically find that a large quantization error brings in extra noise, which influences the convergence and sparsity of the model. Therefore, a threshold quantization strategy with a relatively small error is adopted in QCMD adagrad and QRDA adagrad to improve the signal-to-noise ratio and preserve the sparsity of the model. Both theoretical analyses and empirical results demonstrate the efficacy and efficiency of the proposed algorithms.

Qianggang Ding, Deheng Ye, Tingyang Xu et al.

Graph neural networks (GNNs) have been applied into a variety of graph tasks. Most existing work of GNNs is based on the assumption that the given graph data is optimal, while it is inevitable that there exists missing or incomplete edges in the graph data for training, leading to degraded performance. In this paper, we propose Generative Predictive Network (GPN), a GNN-based joint learning framework that simultaneously learns the graph structure and the downstream task. Specifically, we develop a bilevel optimization framework for this joint learning task, in which the upper optimization (generator) and the lower optimization (predictor) are both instantiated with GNNs. To the best of our knowledge, our method is the first GNN-based bilevel optimization framework for resolving this task. Through extensive experiments, our method outperforms a wide range of baselines using benchmark datasets.

Tiantian Zhang, Kevin Zehua Shen, Zichuan Lin et al.

Replaying past experiences has proven to be a highly effective approach for averting catastrophic forgetting in supervised continual learning. However, some crucial factors are still largely ignored, making it vulnerable to serious failure, when used as a solution to forgetting in continual reinforcement learning, even in the context of perfect memory where all data of previous tasks are accessible in the current task. On the one hand, since most reinforcement learning algorithms are not invariant to the reward scale, the previously well-learned tasks (with high rewards) may appear to be more salient to the current learning process than the current task (with small initial rewards). This causes the agent to concentrate on those salient tasks at the expense of generality on the current task. On the other hand, offline learning on replayed tasks while learning a new task may induce a distributional shift between the dataset and the learned policy on old tasks, resulting in forgetting. In this paper, we introduce RECALL, a replay-enhanced method that greatly improves the plasticity of existing replay-based methods on new tasks while effectively avoiding the recurrence of catastrophic forgetting in continual reinforcement learning. RECALL leverages adaptive normalization on approximate targets and policy distillation on old tasks to enhance generality and stability, respectively. Extensive experiments on the Continual World benchmark show that RECALL performs significantly better than purely perfect memory replay, and achieves comparable or better overall performance against state-of-the-art continual learning methods.

Junyou Li, Qin Zhang, Yangbin Yu et al.

We find that, simply via a sampling-and-voting method, the performance of large language models (LLMs) scales with the number of agents instantiated. Also, this method, termed as Agent Forest, is orthogonal to existing complicated methods to further enhance LLMs, while the degree of enhancement is correlated to the task difficulty. We conduct comprehensive experiments on a wide range of LLM benchmarks to verify the presence of our finding, and to study the properties that can facilitate its occurrence. Our code is publicly available at: https://github.com/MoreAgentsIsAllYouNeed/AgentForest

Fuxiang Zhang, Junyou Li, Yi-Chen Li et al.

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 paper, 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 pre-collected 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.

Haoxuan Pan, Deheng Ye, Xiaoming Duan et al.

We revisit the estimation bias in policy gradients for the discounted episodic Markov decision process (MDP) from Deep Reinforcement Learning (DRL) perspective. The objective is formulated theoretically as the expected returns discounted over the time horizon. One of the major policy gradient biases is the state distribution shift: the state distribution used to estimate the gradients differs from the theoretical formulation in that it does not take into account the discount factor. Existing discussion of the influence of this bias was limited to the tabular and softmax cases in the literature. Therefore, in this paper, we extend it to the DRL setting where the policy is parameterized and demonstrate how this bias can lead to suboptimal policies theoretically. We then discuss why the empirically inaccurate implementations with shifted state distribution can still be effective. We show that, despite such state distribution shift, the policy gradient estimation bias can be reduced in the following three ways: 1) a small learning rate; 2) an adaptive-learning-rate-based optimizer; and 3) KL regularization. Specifically, we show that a smaller learning rate, or, an adaptive learning rate, such as that used by Adam and RSMProp optimizers, makes the policy optimization robust to the bias. We further draw connections between optimizers and the optimization regularization to show that both the KL and the reverse KL regularization can significantly rectify this bias. Moreover, we provide extensive experiments on continuous control tasks to support our analysis. Our paper sheds light on how successful PG algorithms optimize policies in the DRL setting, and contributes insights into the practical issues in DRL.

Yangbin Yu, Qin Zhang, Junyou Li et al.

The emergence of large language models (LLMs) has significantly advanced the simulation of believable interactive agents. However, the substantial cost on maintaining the prolonged agent interactions poses challenge over the deployment of believable LLM-based agents. Therefore, in this paper, we develop Affordable Generative Agents (AGA), a framework for enabling the generation of believable and low-cost interactions on both agent-environment and inter-agents levels. Specifically, for agent-environment interactions, we substitute repetitive LLM inferences with learned policies; while for inter-agent interactions, we model the social relationships between agents and compress auxiliary dialogue information. Extensive experiments on multiple environments show the effectiveness and efficiency of our proposed framework. Also, we delve into the mechanisms of emergent believable behaviors lying in LLM agents, demonstrating that agents can only generate finite behaviors in fixed environments, based upon which, we understand ways to facilitate emergent interaction behaviors. Our code is publicly available at: https://github.com/AffordableGenerativeAgents/Affordable-Generative-Agents.

Zichuan Lin, Xiaokai Huang, Jiate Liu et al.

The estimation of individual treatment effects (ITE) focuses on predicting the outcome changes that result from a change in treatment. A fundamental challenge in observational data is that while we need to infer outcome differences under alternative treatments, we can only observe each individual's outcome under a single treatment. Existing approaches address this limitation either by training with inferred pseudo-outcomes or by creating matched instance pairs. However, recent work has largely overlooked the potential impact of post-treatment variables on the outcome. This oversight prevents existing methods from fully capturing outcome variability, resulting in increased variance in counterfactual predictions. This paper introduces Pseudo-outcome Imputation with Post-treatment Variables for Counterfactual Regression (PIPCFR), a novel approach that incorporates post-treatment variables to improve pseudo-outcome imputation. We analyze the challenges inherent in utilizing post-treatment variables and establish a novel theoretical bound for ITE risk that explicitly connects post-treatment variables to ITE estimation accuracy. Unlike existing methods that ignore these variables or impose restrictive assumptions, PIPCFR learns effective representations that preserve informative components while mitigating bias. Empirical evaluations on both real-world and simulated datasets demonstrate that PIPCFR achieves significantly lower ITE errors compared to existing methods.

Fan Wu, Jiacheng Wei, Ruibo Li et al.

Video-based world models have recently garnered increasing attention for their ability to synthesize diverse and dynamic visual environments. In this paper, we focus on shared world modeling, where a model generates multiple videos from a set of input images, each representing the same underlying world in different camera poses. We propose IC-World, a novel generation framework, enabling parallel generation for all input images via activating the inherent in-context generation capability of large video models. We further finetune IC-World via reinforcement learning, Group Relative Policy Optimization, together with two proposed novel reward models to enforce scene-level geometry consistency and object-level motion consistency among the set of generated videos. Extensive experiments demonstrate that IC-World substantially outperforms state-of-the-art methods in both geometry and motion consistency. To the best of our knowledge, this is the first work to systematically explore the shared world modeling problem with video-based world models.

Fan Wu, Cheng Chen, Zhoujie Fu et al.

Recent diffusion-based text-to-image customization methods have achieved significant success in understanding concrete concepts to control generation processes, such as styles and shapes. However, few efforts dive into the realistic yet challenging customization of physical concepts. The core limitation of current methods arises from the absence of explicitly introducing physical knowledge during training. Even when physics-related words appear in the input text prompts, our experiments consistently demonstrate that these methods fail to accurately reflect the corresponding physical properties in the generated results. In this paper, we propose PhyCustom, a fine-tuning framework comprising two novel regularization losses to activate diffusion model to perform physical customization. Specifically, the proposed isometric loss aims at activating diffusion models to learn physical concepts while decouple loss helps to eliminate the mixture learning of independent concepts. Experiments are conducted on a diverse dataset and our benchmark results demonstrate that PhyCustom outperforms previous state-of-the-art and popular methods in terms of physical customization quantitatively and qualitatively.

Longxiang He, Deheng Ye, Junbo Tan et al.

Pretraining a policy on offline data followed by fine-tuning through online interactions, known as Offline-to-Online Reinforcement Learning (O2O RL), has emerged as a promising paradigm for real-world RL deployment. However, both offline datasets and online interactions in practical environments are often noisy or even maliciously corrupted, severely degrading the performance of O2O RL. Existing works primarily focus on mitigating the conservatism of offline policies via online exploration, while the robustness of O2O RL under data corruption, including states, actions, rewards, and dynamics, is still unexplored. In this work, we observe that data corruption induces heavy-tailed behavior in the policy, thereby substantially degrading the efficiency of online exploration. To address this issue, we incorporate Inverse Probability Weighted (IPW) into the online exploration policy to alleviate heavy-tailedness, and propose a novel, simple yet effective method termed $\textbf{RPEX}$: $\textbf{R}$obust $\textbf{P}$olicy $\textbf{EX}$pansion. Extensive experimental results on D4RL datasets demonstrate that RPEX achieves SOTA O2O performance across a wide range of data corruption scenarios. Code is available at $\href{https://github.com/felix-thu/RPEX}{https://github.com/felix-thu/RPEX}$.

Zheng Zhang, Peilin Zhao, Deheng Ye et al.

Jailbreak attacks aim to exploit large language models (LLMs) by inducing them to generate harmful content, thereby revealing their vulnerabilities. Understanding and addressing these attacks is crucial for advancing the field of LLM safety. Previous jailbreak approaches have mainly focused on direct manipulations of harmful intent, with limited attention to the impact of persona prompts. In this study, we systematically explore the efficacy of persona prompts in compromising LLM defenses. We propose a genetic algorithm-based method that automatically crafts persona prompts to bypass LLM's safety mechanisms. Our experiments reveal that: (1) our evolved persona prompts reduce refusal rates by 50-70% across multiple LLMs, and (2) these prompts demonstrate synergistic effects when combined with existing attack methods, increasing success rates by 10-20%. Our code and data are available at https://github.com/CjangCjengh/Generic_Persona.

Zhihui Xie, Zichuan Lin, Deheng Ye et al.

Recent research in offline reinforcement learning (RL) has demonstrated that return-conditioned supervised learning is a powerful paradigm for decision-making problems. While promising, return conditioning is limited to training data labeled with rewards and therefore faces challenges in learning from unsupervised data. In this work, we aim to utilize generalized future conditioning to enable efficient unsupervised pretraining from reward-free and sub-optimal offline data. We propose Pretrained Decision Transformer (PDT), a conceptually simple approach for unsupervised RL pretraining. PDT leverages future trajectory information as a privileged context to predict actions during training. The ability to make decisions based on both present and future factors enhances PDT's capability for generalization. Besides, this feature can be easily incorporated into a return-conditioned framework for online finetuning, by assigning return values to possible futures and sampling future embeddings based on their respective values. Empirically, PDT outperforms or performs on par with its supervised pretraining counterpart, especially when dealing with sub-optimal data. Further analysis reveals that PDT can extract diverse behaviors from offline data and controllably sample high-return behaviors by online finetuning. Code is available at here.

Zhikang Chen, Sen Cui, Deheng Ye et al.

Large Language Models (LLMs) have demonstrated strong reasoning capabilities through \emph{Chain-of-Thought} (CoT) prompting, which enables step-by-step intermediate reasoning. However, explicit CoT methods rely on discrete token-level reasoning processes that are prone to error propagation and limited by vocabulary expressiveness, often resulting in rigid and inconsistent reasoning trajectories. Recent research has explored implicit or continuous reasoning in latent spaces, allowing models to perform internal reasoning before generating explicit output. Although such approaches alleviate some limitations of discrete CoT, they generally lack explicit mechanisms to enforce consistency among reasoning steps, leading to divergent reasoning paths and unstable outcomes. To address this issue, we propose EBM-CoT, an Energy-Based Chain-of-Thought Calibration framework that refines latent thought representations through an energy-based model (EBM). Our method dynamically adjusts latent reasoning trajectories toward lower-energy, high-consistency regions in the embedding space, improving both reasoning accuracy and consistency without modifying the base language model. Extensive experiments across mathematical, commonsense, and symbolic reasoning benchmarks demonstrate that the proposed framework significantly enhances the consistency and efficiency of multi-step reasoning in LLMs.

Zhifei Xie, Ziyang Ma, Zihang Liu et al.

Reasoning is essential for effective communication and decision-making. While recent advances in LLMs and MLLMs have shown that incorporating explicit reasoning significantly improves understanding and generalization, reasoning in LSMs remains in a nascent stage. Early efforts attempt to transfer the "Thinking-before-Speaking" paradigm from textual models to speech. However, this sequential formulation introduces notable latency, as spoken responses are delayed until reasoning is fully completed, impairing real-time interaction and communication efficiency. To address this, we propose Mini-Omni-Reasoner, a framework that enables reasoning within speech via a novel "Thinking-in-Speaking" formulation. Rather than completing reasoning before producing any verbal output, Mini-Omni-Reasoner interleaves silent reasoning tokens with spoken response tokens at the token level. This design allows continuous speech generation while embedding structured internal reasoning, leveraging the model's high-frequency token processing capability. Although interleaved, local semantic alignment is enforced to ensure that each response token is informed by its preceding reasoning. To support this framework, we introduce Spoken-Math-Problems-3M, a large-scale dataset tailored for interleaved reasoning and response. The dataset ensures that verbal tokens consistently follow relevant reasoning content, enabling accurate and efficient learning of speech-coupled reasoning. Built on a hierarchical Thinker-Talker architecture, Mini-Omni-Reasoner delivers fluent yet logically grounded spoken responses, maintaining both naturalness and precision. On the Spoken-MQA benchmark, it achieves a +19.1% gain in arithmetic reasoning and +6.4% in contextual understanding, with shorter outputs and zero decoding latency.

Tong Wei, Yijun Yang, Junliang Xing et al.

Reinforcement learning with verifiable outcome rewards (RLVR) has effectively scaled up chain-of-thought (CoT) reasoning in large language models (LLMs). Yet, its efficacy in training vision-language model (VLM) agents for goal-directed action reasoning in visual environments is less established. This work investigates this problem through extensive experiments on complex card games, such as 24 points, and embodied tasks from ALFWorld. We find that when rewards are based solely on action outcomes, RL fails to incentivize CoT reasoning in VLMs, instead leading to a phenomenon we termed thought collapse, characterized by a rapid loss of diversity in the agent's thoughts, state-irrelevant and incomplete reasoning, and subsequent invalid actions, resulting in negative rewards. To counteract thought collapse, we highlight the necessity of process guidance and propose an automated corrector that evaluates and refines the agent's reasoning at each RL step. This simple and scalable GTR (Guided Thought Reinforcement) framework trains reasoning and action simultaneously without the need for dense, per-step human labeling. Our experiments demonstrate that GTR significantly enhances the performance and generalization of the LLaVA-7b model across various visual environments, achieving 3-5 times higher task success rates compared to SoTA models with notably smaller model sizes.

Zheng Zhang, Nuoqian Xiao, Qi Chai et al.

Large Language Model (LLM) agents have demonstrated impressive capabilities in social deduction games (SDGs) like Werewolf, where strategic reasoning and social deception are essential. However, current approaches remain limited to textual information, ignoring crucial multimodal cues such as facial expressions and tone of voice that humans naturally use to communicate. Moreover, existing SDG agents primarily focus on inferring other players' identities without modeling how others perceive themselves or fellow players. To address these limitations, we use One Night Ultimate Werewolf (ONUW) as a testbed and present MultiMind, the first framework integrating multimodal information into SDG agents. MultiMind processes facial expressions and vocal tones alongside verbal content, while employing a Theory of Mind (ToM) model to represent each player's suspicion levels toward others. By combining this ToM model with Monte Carlo Tree Search (MCTS), our agent identifies communication strategies that minimize suspicion directed at itself. Through comprehensive evaluation in both agent-versus-agent simulations and studies with human players, we demonstrate MultiMind's superior performance in gameplay. Our work presents a significant advancement toward LLM agents capable of human-like social reasoning across multimodal domains.

Kanefumi Matsuyama, Kefan Su, Jiangxing Wang et al.

Cooperative multi-agent reinforcement learning (MARL) aims to develop agents that can collaborate effectively. However, most cooperative MARL methods overfit training agents, making learned policies not generalize well to unseen collaborators, which is a critical issue for real-world deployment. Some methods attempt to address the generalization problem but require prior knowledge or predefined policies of new teammates, limiting real-world applications. To this end, we propose a hierarchical MARL approach to enable generalizable cooperation via role diversity, namely CORD. CORD's high-level controller assigns roles to low-level agents by maximizing the role entropy with constraints. We show this constrained objective can be decomposed into causal influence in role that enables reasonable role assignment, and role heterogeneity that yields coherent, non-redundant role clusters. Evaluated on a variety of cooperative multi-agent tasks, CORD achieves better performance than baselines, especially in generalization tests. Ablation studies further demonstrate the efficacy of the constrained objective in generalizable cooperation.

Yao Tang, Zhihui Xie, Zichuan Lin et al.

Masked prediction has emerged as a promising pretraining paradigm in offline reinforcement learning (RL) due to its versatile masking schemes, enabling flexible inference across various downstream tasks with a unified model. Despite the versatility of masked prediction, it remains unclear how to balance the learning of skills at different levels of complexity. To address this, we propose CurrMask, a curriculum masking pretraining paradigm for sequential decision making. Motivated by how humans learn by organizing knowledge in a curriculum, CurrMask adjusts its masking scheme during pretraining for learning versatile skills. Through extensive experiments, we show that CurrMask exhibits superior zero-shot performance on skill prompting tasks, goal-conditioned planning tasks, and competitive finetuning performance on offline RL tasks. Additionally, our analysis of training dynamics reveals that CurrMask gradually acquires skills of varying complexity by dynamically adjusting its masking scheme.

Zheng Zhang, Jiarui He, Yuchen Cai et al.

As large language model (LLM) agents increasingly automate complex web tasks, they boost productivity while simultaneously introducing new security risks. However, relevant studies on web agent attacks remain limited. Existing red-teaming approaches mainly rely on manually crafted attack strategies or static models trained offline. Such methods fail to capture the underlying behavioral patterns of web agents, making it difficult to generalize across diverse environments. In web agent attacks, success requires the continuous discovery and evolution of attack strategies. To this end, we propose Genesis, a novel agentic framework composed of three modules: Attacker, Scorer, and Strategist. The Attacker generates adversarial injections by integrating the genetic algorithm with a hybrid strategy representation. The Scorer evaluates the target web agent's responses to provide feedback. The Strategist dynamically uncovers effective strategies from interaction logs and compiles them into a continuously growing strategy library, which is then re-deployed to enhance the Attacker's effectiveness. Extensive experiments across various web tasks show that our framework discovers novel strategies and consistently outperforms existing attack baselines.

Zhang Zheng, Deheng Ye, Peilin Zhao et al.

Large language model (LLM) agents have shown remarkable progress in social deduction games (SDGs). However, existing approaches primarily focus on information processing and strategy selection, overlooking the significance of persuasive communication in influencing other players' beliefs and responses. In SDGs, success depends not only on making correct deductions but on convincing others to response in alignment with one's intent. To address this limitation, we formalize turn-based dialogue in SDGs as a Stackelberg competition, where the current player acts as the leader who strategically influences the follower's response. Building on this theoretical foundation, we propose a reinforcement learning framework that trains agents to optimize utterances for persuasive impact. Through comprehensive experiments across three diverse SDGs, we demonstrate that our agents significantly outperform baselines. This work represents a significant step toward developing AI agents capable of strategic social influence, with implications extending to scenarios requiring persuasive communication.

He Zhao, Zhiwei Zeng, Yongwei Wang et al.

Heterogeneous Graph Neural Networks (HGNNs) are increasingly recognized for their performance in areas like the web and e-commerce, where resilience against adversarial attacks is crucial. However, existing adversarial attack methods, which are primarily designed for homogeneous graphs, fall short when applied to HGNNs due to their limited ability to address the structural and semantic complexity of HGNNs. This paper introduces HGAttack, the first dedicated gray box evasion attack method for heterogeneous graphs. We design a novel surrogate model to closely resemble the behaviors of the target HGNN and utilize gradient-based methods for perturbation generation. Specifically, the proposed surrogate model effectively leverages heterogeneous information by extracting meta-path induced subgraphs and applying GNNs to learn node embeddings with distinct semantics from each subgraph. This approach improves the transferability of generated attacks on the target HGNN and significantly reduces memory costs. For perturbation generation, we introduce a semantics-aware mechanism that leverages subgraph gradient information to autonomously identify vulnerable edges across a wide range of relations within a constrained perturbation budget. We validate HGAttack's efficacy with comprehensive experiments on three datasets, providing empirical analyses of its generated perturbations. Outperforming baseline methods, HGAttack demonstrated significant efficacy in diminishing the performance of target HGNN models, affirming the effectiveness of our approach in evaluating the robustness of HGNNs against adversarial attacks.

Zifan Wu, Chao Yu, Deheng Ye et al.

We present Coordinated Proximal Policy Optimization (CoPPO), an algorithm that extends the original Proximal Policy Optimization (PPO) to the multi-agent setting. The key idea lies in the coordinated adaptation of step size during the policy update process among multiple agents. We prove the monotonicity of policy improvement when optimizing a theoretically-grounded joint objective, and derive a simplified optimization objective based on a set of approximations. We then interpret that such an objective in CoPPO can achieve dynamic credit assignment among agents, thereby alleviating the high variance issue during the concurrent update of agent policies. Finally, we demonstrate that CoPPO outperforms several strong baselines and is competitive with the latest multi-agent PPO method (i.e. MAPPO) under typical multi-agent settings, including cooperative matrix games and the StarCraft II micromanagement tasks.

Yiming Gao, Bei Shi, Xueying Du et al.

Recently, many researchers have made successful progress in building the AI systems for MOBA-game-playing with deep reinforcement learning, such as on Dota 2 and Honor of Kings. Even though these AI systems have achieved or even exceeded human-level performance, they still suffer from the lack of policy diversity. In this paper, we propose a novel Macro-Goals Guided framework, called MGG, to learn diverse policies in MOBA games. MGG abstracts strategies as macro-goals from human demonstrations and trains a Meta-Controller to predict these macro-goals. To enhance policy diversity, MGG samples macro-goals from the Meta-Controller prediction and guides the training process towards these goals. Experimental results on the typical MOBA game Honor of Kings demonstrate that MGG can execute diverse policies in different matches and lineups, and also outperform the state-of-the-art methods over 102 heroes.

Shiyu Huang, Wenze Chen, Longfei Zhang et al.

Deep reinforcement learning (DRL) has achieved super-human performance on complex video games (e.g., StarCraft II and Dota II). However, current DRL systems still suffer from challenges of multi-agent coordination, sparse rewards, stochastic environments, etc. In seeking to address these challenges, we employ a football video game, e.g., Google Research Football (GRF), as our testbed and develop an end-to-end learning-based AI system (denoted as TiKick) to complete this challenging task. In this work, we first generated a large replay dataset from the self-playing of single-agent experts, which are obtained from league training. We then developed a distributed learning system and new offline algorithms to learn a powerful multi-agent AI from the fixed single-agent dataset. To the best of our knowledge, Tikick is the first learning-based AI system that can take over the multi-agent Google Research Football full game, while previous work could either control a single agent or experiment on toy academic scenarios. Extensive experiments further show that our pre-trained model can accelerate the training process of the modern multi-agent algorithm and our method achieves state-of-the-art performances on various academic scenarios.

Hua Wei, Deheng Ye, Zhao Liu et al.

Offline reinforcement learning (RL) tries to learn the near-optimal policy with recorded offline experience without online exploration. Current offline RL research includes: 1) generative modeling, i.e., approximating a policy using fixed data; and 2) learning the state-action value function. While most research focuses on the state-action function part through reducing the bootstrapping error in value function approximation induced by the distribution shift of training data, the effects of error propagation in generative modeling have been neglected. In this paper, we analyze the error in generative modeling. We propose AQL (action-conditioned Q-learning), a residual generative model to reduce policy approximation error for offline RL. We show that our method can learn more accurate policy approximations in different benchmark datasets. In addition, we show that the proposed offline RL method can learn more competitive AI agents in complex control tasks under the multiplayer online battle arena (MOBA) game Honor of Kings.

Menghui Zhu, Minghuan Liu, Jian Shen et al.

In Goal-oriented Reinforcement learning, relabeling the raw goals in past experience to provide agents with hindsight ability is a major solution to the reward sparsity problem. In this paper, to enhance the diversity of relabeled goals, we develop FGI (Foresight Goal Inference), a new relabeling strategy that relabels the goals by looking into the future with a learned dynamics model. Besides, to improve sample efficiency, we propose to use the dynamics model to generate simulated trajectories for policy training. By integrating these two improvements, we introduce the MapGo framework (Model-Assisted Policy Optimization for Goal-oriented tasks). In our experiments, we first show the effectiveness of the FGI strategy compared with the hindsight one, and then show that the MapGo framework achieves higher sample efficiency when compared to model-free baselines on a set of complicated tasks.

Sheng Chen, Menghui Zhu, Deheng Ye et al.

Hero drafting is essential in MOBA game playing as it builds the team of each side and directly affects the match outcome. State-of-the-art drafting methods fail to consider: 1) drafting efficiency when the hero pool is expanded; 2) the multi-round nature of a MOBA 5v5 match series, i.e., two teams play best-of-N and the same hero is only allowed to be drafted once throughout the series. In this paper, we formulate the drafting process as a multi-round combinatorial game and propose a novel drafting algorithm based on neural networks and Monte-Carlo tree search, named JueWuDraft. Specifically, we design a long-term value estimation mechanism to handle the best-of-N drafting case. Taking Honor of Kings, one of the most popular MOBA games at present, as a running case, we demonstrate the practicality and effectiveness of JueWuDraft when compared to state-of-the-art drafting methods.

Deheng Ye, Guibin Chen, Wen Zhang et al.

MOBA games, e.g., Honor of Kings, League of Legends, and Dota 2, pose grand challenges to AI systems such as multi-agent, enormous state-action space, complex action control, etc. Developing AI for playing MOBA games has raised much attention accordingly. However, existing work falls short in handling the raw game complexity caused by the explosion of agent combinations, i.e., lineups, when expanding the hero pool in case that OpenAI's Dota AI limits the play to a pool of only 17 heroes. As a result, full MOBA games without restrictions are far from being mastered by any existing AI system. In this paper, we propose a MOBA AI learning paradigm that methodologically enables playing full MOBA games with deep reinforcement learning. Specifically, we develop a combination of novel and existing learning techniques, including curriculum self-play learning, policy distillation, off-policy adaption, multi-head value estimation, and Monte-Carlo tree-search, in training and playing a large pool of heroes, meanwhile addressing the scalability issue skillfully. Tested on Honor of Kings, a popular MOBA game, we show how to build superhuman AI agents that can defeat top esports players. The superiority of our AI is demonstrated by the first large-scale performance test of MOBA AI agent in the literature.

Deheng Ye, Guibin Chen, Peilin Zhao et al.

We present JueWu-SL, the first supervised-learning-based artificial intelligence (AI) program that achieves human-level performance in playing multiplayer online battle arena (MOBA) games. Unlike prior attempts, we integrate the macro-strategy and the micromanagement of MOBA-game-playing into neural networks in a supervised and end-to-end manner. Tested on Honor of Kings, the most popular MOBA at present, our AI performs competitively at the level of High King players in standard 5v5 games.

Deheng Ye, Zhao Liu, Mingfei Sun et al.

We study the reinforcement learning problem of complex action control in the Multi-player Online Battle Arena (MOBA) 1v1 games. This problem involves far more complicated state and action spaces than those of traditional 1v1 games, such as Go and Atari series, which makes it very difficult to search any policies with human-level performance. In this paper, we present a deep reinforcement learning framework to tackle this problem from the perspectives of both system and algorithm. Our system is of low coupling and high scalability, which enables efficient explorations at large scale. Our algorithm includes several novel strategies, including control dependency decoupling, action mask, target attention, and dual-clip PPO, with which our proposed actor-critic network can be effectively trained in our system. Tested on the MOBA game Honor of Kings, our AI agent, called Tencent Solo, can defeat top professional human players in full 1v1 games.