Zhihan Liu

Zhenghai Xue, Zhenghao Peng, Quanyi Li et al.

The Teacher-Student Framework (TSF) is a reinforcement learning setting where a teacher agent guards the training of a student agent by intervening and providing online demonstrations. Assuming optimal, the teacher policy has the perfect timing and capability to intervene in the learning process of the student agent, providing safety guarantee and exploration guidance. Nevertheless, in many real-world settings it is expensive or even impossible to obtain a well-performing teacher policy. In this work, we relax the assumption of a well-performing teacher and develop a new method that can incorporate arbitrary teacher policies with modest or inferior performance. We instantiate an Off-Policy Reinforcement Learning algorithm, termed Teacher-Student Shared Control (TS2C), which incorporates teacher intervention based on trajectory-based value estimation. Theoretical analysis validates that the proposed TS2C algorithm attains efficient exploration and substantial safety guarantee without being affected by the teacher's own performance. Experiments on various continuous control tasks show that our method can exploit teacher policies at different performance levels while maintaining a low training cost. Moreover, the student policy surpasses the imperfect teacher policy in terms of higher accumulated reward in held-out testing environments. Code is available at https://metadriverse.github.io/TS2C.

Zhihan Liu, Hao Hu, Shenao Zhang et al.

Large language models (LLMs) demonstrate impressive reasoning abilities, but translating reasoning into actions in the real world remains challenging. In particular, it remains unclear how to complete a given task provably within a minimum number of interactions with the external environment, e.g., through an internal mechanism of reasoning. To this end, we propose a principled framework with provable regret guarantees to orchestrate reasoning and acting, which we call "reason for future, act for now" (\texttt{RAFA}). Specifically, we design a prompt template for reasoning that learns from the memory buffer and plans a future trajectory over a long horizon ("reason for future"). At each step, the LLM agent takes the initial action of the planned trajectory ("act for now"), stores the collected feedback in the memory buffer, and reinvokes the reasoning routine to replan the future trajectory from the new state. The key idea is to cast reasoning in LLMs as learning and planning in Bayesian adaptive Markov decision processes (MDPs). Correspondingly, we prompt LLMs to form an updated posterior of the unknown environment from the memory buffer (learning) and generate an optimal trajectory for multiple future steps that maximizes a value function (planning). The learning and planning subroutines are performed in an "in-context" manner to emulate the actor-critic update for MDPs. Our theoretical analysis proves that the novel combination of long-term reasoning and short-term acting achieves a $\sqrt{T}$ regret. Here, $T$ denotes the number of online interactions. In particular, the regret bound highlights an intriguing interplay between the prior knowledge obtained through pretraining and the uncertainty reduction achieved by reasoning and acting. Our empirical validation shows that it outperforms various existing frameworks and achieves nearly perfect scores on a few benchmarks.

Nuoya Xiong, Zhihan Liu, Zhaoran Wang et al.

We study multi-agent reinforcement learning (MARL) for the general-sum Markov Games (MGs) under the general function approximation. In order to find the minimum assumption for sample-efficient learning, we introduce a novel complexity measure called the Multi-Agent Decoupling Coefficient (MADC) for general-sum MGs. Using this measure, we propose the first unified algorithmic framework that ensures sample efficiency in learning Nash Equilibrium, Coarse Correlated Equilibrium, and Correlated Equilibrium for both model-based and model-free MARL problems with low MADC. We also show that our algorithm provides comparable sublinear regret to the existing works. Moreover, our algorithm combines an equilibrium-solving oracle with a single objective optimization subprocedure that solves for the regularized payoff of each deterministic joint policy, which avoids solving constrained optimization problems within data-dependent constraints (Jin et al. 2020; Wang et al. 2023) or executing sampling procedures with complex multi-objective optimization problems (Foster et al. 2023), thus being more amenable to empirical implementation.

Saizhuo Wang, Zhihan Liu, Zhaoran Wang et al.

Large Language Models (LLMs) are versatile, yet they often falter in tasks requiring deep and reliable reasoning due to issues like hallucinations, limiting their applicability in critical scenarios. This paper introduces a rigorously designed framework for creating LLMs that effectively anchor knowledge and employ a closed-loop reasoning process, enhancing their capability for in-depth analysis. We dissect the framework to illustrate the contribution of each component to the LLMs' performance, offering a theoretical assurance of improved reasoning under well-defined assumptions.

Yuanbo Xie, Yingjie Zhang, Yulin Li et al.

Retrieval-Augmented Generation (RAG) systems augment large language models with external knowledge, yet introduce a critical security vulnerability: RAG Knowledge Base Leakage, wherein adversarial prompts can induce the model to divulge retrieved proprietary content. Recent studies reveal that such leakage can be executed through adaptive and iterative attack strategies (named RAG extraction attack), while effective countermeasures remain notably lacking. To bridge this gap, we propose CanaryRAG, a runtime defense mechanism inspired by stack canaries in software security. CanaryRAG embeds carefully designed canary tokens into retrieved chunks and reformulates RAG extraction defense as a dual-path runtime integrity game. Leakage is detected in real time whenever either the target or oracle path violates its expected canary behavior, including under adaptive suppression and obfuscation. Extensive evaluations against existing attacks demonstrate that CanaryRAG provides robust defense, achieving substantially lower chunk recovery rates than state-of-the-art baselines while imposing negligible impact on task performance and inference latency. Moreover, as a plug-and-play solution, CanaryRAG can be seamlessly integrated into arbitrary RAG pipelines without requiring retraining or structural modifications, offering a practical and scalable safeguard for proprietary data.

Shenao Zhang, Sirui Zheng, Shuqi Ke et al.

Reinforcement learning (RL) has become the de facto standard practice for sequential decision-making problems by improving future acting policies with feedback. However, RL algorithms may require extensive trial-and-error interactions to collect useful feedback for improvement. On the other hand, recent developments in large language models (LLMs) have showcased impressive capabilities in language understanding and generation, yet they fall short in exploration and self-improvement capabilities for planning tasks, lacking the ability to autonomously refine their responses based on feedback. Therefore, in this paper, we study how the policy prior provided by the LLM can enhance the sample efficiency of RL algorithms. Specifically, we develop an algorithm named LINVIT that incorporates LLM guidance as a regularization factor in value-based RL, leading to significant reductions in the amount of data needed for learning, particularly when the difference between the ideal policy and the LLM-informed policy is small, which suggests that the initial policy is close to optimal, reducing the need for further exploration. Additionally, we present a practical algorithm SLINVIT that simplifies the construction of the value function and employs subgoals to reduce the search complexity. Our experiments across three interactive environments ALFWorld, InterCode, and BlocksWorld demonstrate that our method achieves state-of-the-art success rates and also surpasses previous RL and LLM approaches in terms of sample efficiency. Our code is available at https://github.com/agentification/Language-Integrated-VI.

Zhihan Liu, Lin Guan, Yixin Nie et al.

Generalist LLM agents are often post-trained on a narrow set of environments but deployed across far broader, unseen domains. In this work, we investigate the challenge of agentic post-training when the eventual test domains are unknown. Specifically, we analyze which properties of reinforcement learning (RL) environments and modeling choices have the greatest influence on out-of-domain performance. First, we identify two environment axes that strongly correlate with cross-domain generalization: (i) state information richness, i.e., the amount of information for the agent to process from the state, and (ii) planning complexity, estimated via goal reachability and trajectory length under a base policy. Notably, domain realism and text-level similarity are not the primary factors; for instance, the simple grid-world domain Sokoban leads to even stronger generalization in SciWorld than the more realistic ALFWorld. Motivated by these findings, we further show that increasing state information richness alone can already effectively improve cross-domain robustness. We propose a randomization technique, which is low-overhead and broadly applicable: add small amounts of distractive goal-irrelevant features to the state to make it richer without altering the task. Beyond environment-side properties, we also examine several modeling choices: (a) SFT warmup or mid-training helps prevent catastrophic forgetting during RL but undermines generalization to domains that are not included in the mid-training datamix; and (b) turning on step-by-step thinking during RL, while not always improving in-domain performance, plays a crucial role in preserving generalization.

Ruijie Xu, Zhihan Liu, Yongfei Liu et al.

We address the challenge of online Reinforcement Learning from Human Feedback (RLHF) with a focus on self-rewarding alignment methods. In online RLHF, obtaining feedback requires interaction with the environment, which can be costly when using additional reward models or the GPT-4 API. Current self-rewarding approaches rely heavily on the discriminator's judgment capabilities, which are effective for large-scale models but challenging to transfer to smaller ones. To address these limitations, we propose a novel, only-prompting self-rewarding online algorithm that generates preference datasets without relying on judgment capabilities. Additionally, we employ fine-grained arithmetic control over the optimality gap between positive and negative examples, generating more hard negatives in the later stages of training to help the model better capture subtle human preferences. Finally, we conduct extensive experiments on two base models, Mistral-7B and Mistral-Instruct-7B, which significantly bootstrap the performance of the reference model, achieving 34.5% in the Length-controlled Win Rates of AlpacaEval 2.0.

Zhihan Liu, Yubo Chai, Jianfeng Li

The advent of Large Language Models (LLMs) has created new opportunities for the automation of scientific research spanning both experimental processes and computational simulations. This study explores the feasibility of constructing an autonomous simulation agent (ASA) powered by LLMs through prompt engineering and automated program design to automate the entire simulation research process according to a human-provided research plan. This process includes experimental design, remote upload and simulation execution, data analysis, and report compilation. Using a well-studied simulation problem of polymer chain conformations as a test case, we assessed the long-task completion and reliability of ASAs powered by different LLMs, including GPT-4o, Claude-3.5, etc. Our findings revealed that ASA-GPT-4o achieved near-flawless execution on designated research missions, underscoring the potential of methods like ASA to achieve automation in simulation research processes to enhance research efficiency. The outlined automation can be iteratively performed for up to 20 cycles without human intervention, illustrating the potential of ASA for long-task workflow automation. Additionally, we discussed the intrinsic traits of ASA in managing extensive tasks, focusing on self-validation mechanisms, and the balance between local attention and global oversight.

Hongyi Guo, Zhihan Liu, Yufeng Zhang et al.

Large Language Models (LLMs) harness extensive data from the Internet, storing a broad spectrum of prior knowledge. While LLMs have proven beneficial as decision-making aids, their reliability is hampered by limitations in reasoning, hallucination phenomenon, and so on. On the other hand, Monte-Carlo Tree Search (MCTS) is a heuristic search algorithm that provides reliable decision-making solutions, achieved through recursive rollouts and self-play. However, the effectiveness of MCTS relies heavily on heuristic pruning and external value functions, particularly in complex decision scenarios. This work introduces an innovative approach that bolsters LLMs with MCTS self-play to efficiently resolve deterministic turn-based zero-sum games (DTZG), such as chess and go, without the need for additional training. Specifically, we utilize LLMs as both action pruners and proxies for value functions without the need for additional training. We theoretically prove that the suboptimality of the estimated value in our proposed method scales with $\tilde{\mathcal O}\Bigl(\frac{|\tilde {\mathcal A}|}{\sqrt{N}} + ε_\mathrm{pruner} + ε_\mathrm{critic}\Bigr)$, where \(N\) is the number of simulations, $|\tilde {\mathcal A}|$ is the cardinality of the pruned action space by LLM, and $ε_\mathrm{pruner}$ and $ε_\mathrm{critic}$ quantify the errors incurred by adopting LLMs as action space pruner and value function proxy, respectively. Our experiments in chess and go demonstrate the capability of our method to address challenges beyond the scope of MCTS and improve the performance of the directly application of LLMs.

Han Zhong, Yutong Yin, Shenao Zhang et al.

Large Language Models (LLMs) have demonstrated remarkable capabilities in complex reasoning tasks, yet generating reliable reasoning processes remains a significant challenge. We present a unified probabilistic framework that formalizes LLM reasoning through a novel graphical model incorporating latent thinking processes and evaluation signals. Within this framework, we introduce the Bootstrapping Reinforced Thinking Process (BRiTE) algorithm, which works in two steps. First, it generates high-quality rationales by approximating the optimal thinking process through reinforcement learning, using a novel reward shaping mechanism. Second, it enhances the base LLM by maximizing the joint probability of rationale generation with respect to the model's parameters. Theoretically, we demonstrate BRiTE's convergence at a rate of $1/T$ with $T$ representing the number of iterations. Empirical evaluations on math and coding benchmarks demonstrate that our approach consistently improves performance across different base models without requiring human-annotated thinking processes. In addition, BRiTE demonstrates superior performance compared to existing algorithms that bootstrap thinking processes use alternative methods such as rejection sampling, and can even match or exceed the results achieved through supervised fine-tuning with human-annotated data.

Zhihan Liu, Shenao Zhang, Yongfei Liu et al.

Direct preference learning offers a promising and computation-efficient beyond supervised fine-tuning (SFT) for improving code generation in coding large language models (LMs). However, the scarcity of reliable preference data is a bottleneck for the performance of direct preference learning to improve the coding accuracy of code LMs. In this paper, we introduce \underline{\textbf{D}}irect Preference Learning with Only \underline{\textbf{S}}elf-Generated \underline{\textbf{T}}ests and \underline{\textbf{C}}ode (DSTC), a framework that leverages only self-generated code snippets and tests to construct reliable preference pairs such that direct preference learning can improve LM coding accuracy without external annotations. DSTC combines a minimax selection process and test-code concatenation to improve preference pair quality, reducing the influence of incorrect self-generated tests and enhancing model performance without the need for costly reward models. When applied with direct preference learning methods such as Direct Preference Optimization (DPO) and Kahneman-Tversky Optimization (KTO), DSTC yields stable improvements in coding accuracy (pass@1 score) across diverse coding benchmarks, including HumanEval, MBPP, and BigCodeBench, demonstrating both its effectiveness and scalability for models of various sizes. This approach autonomously enhances code generation accuracy across LLMs of varying sizes, reducing reliance on expensive annotated coding datasets.

Kai Zhang, Xiangchao Chen, Bo Liu et al. · microsoft-research

A long-term goal of language agents is to learn and improve through their own experience, ultimately outperforming humans in complex, real-world tasks. However, training agents from experience data with reinforcement learning remains difficult in many environments, which either lack verifiable rewards (e.g., websites) or require inefficient long-horizon rollouts (e.g., multi-turn tool use). As a result, most current agents rely on supervised fine-tuning on expert data, which is challenging to scale and generalizes poorly. This limitation stems from the nature of expert demonstrations: they capture only a narrow range of scenarios and expose the agent to limited environment diversity. We address this limitation with a middle-ground paradigm we call early experience: interaction data generated by the agent's own actions, where the resulting future states serve as supervision without reward signals. Within this paradigm we study two strategies of using such data: (1) Implicit world modeling, which uses collected states to ground the policy in environment dynamics; and (2) Self-reflection, where the agent learns from its suboptimal actions to improve reasoning and decision-making. We evaluate across eight diverse environments and multiple model families. Our approaches consistently improve effectiveness and out-of-domain generalization, highlighting the value of early experience. Moreover, in environments with verifiable rewards, our results provide promising signals that early experience offers a strong foundation for subsequent reinforcement learning, positioning it as a practical bridge between imitation learning and fully experience-driven agents.

Yuxiao Yang, Shenao Zhang, Zhihan Liu et al.

This work focuses on building a task planner for Embodied Instruction Following (EIF) using Large Language Models (LLMs). Previous works typically train a planner to imitate expert trajectories, treating this as a supervised task. While these methods achieve competitive performance, they often lack sufficient robustness. When a suboptimal action is taken, the planner may encounter an out-of-distribution state, which can lead to task failure. In contrast, we frame the task as a Partially Observable Markov Decision Process (POMDP) and aim to develop a robust planner under a few-shot assumption. Thus, we propose a closed-loop planner with an adaptation module and a novel hindsight method, aiming to use as much information as possible to assist the planner. Our experiments on the ALFRED dataset indicate that our planner achieves competitive performance under a few-shot assumption. For the first time, our few-shot agent's performance approaches and even surpasses that of the full-shot supervised agent.

Rui Zheng, Hongyi Guo, Zhihan Liu et al.

The standard Reinforcement Learning from Human Feedback (RLHF) framework primarily focuses on optimizing the performance of large language models using pre-collected prompts. However, collecting prompts that provide comprehensive coverage is both tedious and challenging, and often fails to include scenarios that LLMs need to improve on the most. In this paper, we investigate alignment through the lens of two-agent games, involving iterative interactions between an adversarial and a defensive agent. The adversarial agent's task at each step is to generate prompts that expose the weakness of the defensive agent. In return, the defensive agent seeks to improve its responses to these newly identified prompts it struggled with, based on feedback from the reward model. We theoretically demonstrate that this iterative reinforcement learning optimization converges to a Nash Equilibrium for the game induced by the agents. Experimental results in safety scenarios demonstrate that learning in such a competitive environment not only fully trains agents but also leads to policies with enhanced generalization capabilities for both adversarial and defensive agents.

Zhihan Liu, Miao Lu, Wei Xiong et al.

In online reinforcement learning (online RL), balancing exploration and exploitation is crucial for finding an optimal policy in a sample-efficient way. To achieve this, existing sample-efficient online RL algorithms typically consist of three components: estimation, planning, and exploration. However, in order to cope with general function approximators, most of them involve impractical algorithmic components to incentivize exploration, such as optimization within data-dependent level-sets or complicated sampling procedures. To address this challenge, we propose an easy-to-implement RL framework called \textit{Maximize to Explore} (\texttt{MEX}), which only needs to optimize \emph{unconstrainedly} a single objective that integrates the estimation and planning components while balancing exploration and exploitation automatically. Theoretically, we prove that \texttt{MEX} achieves a sublinear regret with general function approximations for Markov decision processes (MDP) and is further extendable to two-player zero-sum Markov games (MG). Meanwhile, we adapt deep RL baselines to design practical versions of \texttt{MEX}, in both model-free and model-based manners, which can outperform baselines by a stable margin in various MuJoCo environments with sparse rewards. Compared with existing sample-efficient online RL algorithms with general function approximations, \texttt{MEX} achieves similar sample efficiency while enjoying a lower computational cost and is more compatible with modern deep RL methods.

Zhihan Liu, Yufeng Zhang, Zuyue Fu et al.

In generative adversarial imitation learning (GAIL), the agent aims to learn a policy from an expert demonstration so that its performance cannot be discriminated from the expert policy on a certain predefined reward set. In this paper, we study GAIL in both online and offline settings with linear function approximation, where both the transition and reward function are linear in the feature maps. Besides the expert demonstration, in the online setting the agent can interact with the environment, while in the offline setting the agent only accesses an additional dataset collected by a prior. For online GAIL, we propose an optimistic generative adversarial policy optimization algorithm (OGAP) and prove that OGAP achieves $\widetilde{\mathcal{O}}(H^2 d^{3/2}K^{1/2}+KH^{3/2}dN_1^{-1/2})$ regret. Here $N_1$ represents the number of trajectories of the expert demonstration, $d$ is the feature dimension, and $K$ is the number of episodes. For offline GAIL, we propose a pessimistic generative adversarial policy optimization algorithm (PGAP). For an arbitrary additional dataset, we obtain the optimality gap of PGAP, achieving the minimax lower bound in the utilization of the additional dataset. Assuming sufficient coverage on the additional dataset, we show that PGAP achieves $\widetilde{\mathcal{O}}(H^{2}dK^{-1/2} +H^2d^{3/2}N_2^{-1/2}+H^{3/2}dN_1^{-1/2} \ )$ optimality gap. Here $N_2$ represents the number of trajectories of the additional dataset with sufficient coverage.