arXiv subjects

Siliang Zeng

Publications and source records attributed to Siliang Zeng.

15 recordsLinked to original sources

arXiv2026

Stabilizing Off-Policy Training for Long-Horizon LLM Agent via Turn-Level Importance Sampling and Clipping-Triggered Normalization

Reinforcement learning (RL) algorithms such as PPO and GRPO are widely used to train large language models (LLMs) for multi-turn agentic tasks. However, in off-policy training pipelines, these methods often exhibit unstable optimization dynamics and are prone to performance collapse. Through empirical analysis, we identify two fundamental sources of instability in this setting: (1)~a granularity mismatch between token-level policy optimization and turn-structured interactions, and (2) high-variance and unreliable gradient updates induced by off-policy importance sampling and inaccurate advantage estimation. To address these challenges, we propose SORL, \underline{S}tabilizing \underline{O}ff-Policy \underline{R}einforcement \underline{L}earning for Long-Horizon Agent Training. SORL introduces principled mechanisms that align policy optimization with the structure of multi-turn interactions and adaptively suppress unreliable off-policy updates, yielding more conservative and robust learning dynamics. Within this framework, we instantiate two stabilized algorithms: SO-PPO and SO-GRPO. Both algorithms are designed to mitigate gradient variance and prevent optimization collapse without requiring careful early stopping or heuristic tuning. We evaluate SO-PPO and SO-GRPO on a range of multi-turn search benchmarks, including general question answering, multi-hop question answering, and medical multiple-choice QA tasks. Experimental results show that both methods consistently prevent training instabilities and performance collapses observed in standard PPO and GRPO, maintain lower clipping ratios and more stable optimization trajectories, and achieve superior or comparable task performance. These results demonstrate that the proposed algorithm provides a practical, scalable, and general framework for stabilizing reinforcement learning in multi-turn LLM agent training.

cs.LG↗

arXiv2025

Aligning Frozen LLMs by Reinforcement Learning: An Iterative Reweight-then-Optimize Approach

Aligning large language models (LLMs) with human preferences usually requires fine-tuning methods such as RLHF and DPO. These methods directly optimize the model parameters, so they cannot be used in test-time to improve model performance, nor are they applicable when the model weights are not accessible. In contrast, test-time methods sidestep weight updates by leveraging reward functions to guide and improve output quality. However, they incur high inference costs, and their one-shot guidance is often based on imperfect reward or value functions, leading to suboptimal outputs. In this work, we present a method named Iterative Reweight-then-Optimize (IRO), a reinforcement learning (RL) framework that performs RL-style alignment of the (frozen) base model without touching its parameters. During training, each iteration (i) samples candidates from the base model, (ii) resamples using current value functions, and (iii) trains a new lightweight value function that guides the next decoding pass. At test time, the value functions are used to guide the base model generation via a search-based optimization process. Notably, users can apply IRO to align a model on their own dataset, similar to OpenAI's reinforcement fine-tuning (RFT), but without requiring access to the model weights.

cs.LG↗

arXiv2025

Reinforcing Multi-Turn Reasoning in LLM Agents via Fine-Grained Reward Structure and Credit Assignment

Reinforcement Learning (RL) approaches have been wildly used to enhance the reasoning capabilities of Large Language Model (LLM) agents in long-horizon, multi-turn scenarios. Such interactions can be formalized as turn-level Markov decision processes (MDPs), where intermediate rewards are often available. However, most prior work relies on sparse trajectory-level rewards, resulting in poor credit assignment, while dense turn-level rewards remain underexplored. In this paper, we investigate how to effectively leverage dense turn-level reward structures in RL algorithms, specifically Group Relative Policy Optimization (GRPO) and Proximal Policy Optimization (PPO), to enable fine-grained credit assignment. We categorize reward structures into three types based on their granularity: (1) terminal reward; (2) delayed reward; (3) per-turn reward, each corresponding to a distinct turn-level MDP formulation, and derive GRPO and PPO algorithms tailored to each case, respectively. Experiments on multi-turn search and game agent tasks show that, for both GRPO and PPO, dense per-turn reward structures consistently outperform sparse terminal and delayed reward structures in terms of training dynamics and numerical results. Furthermore, on search tasks, PPO with dense per-turn rewards achieves greater training stability and faster convergence, and delivers the highest answer correctness among all state-of-the-art methods across diverse question-answering datasets.

cs.LG↗

arXiv2025

Understanding Inverse Reinforcement Learning under Overparameterization: Non-Asymptotic Analysis and Global Optimality

The goal of the Inverse reinforcement learning (IRL) task is to identify the underlying reward function and the corresponding optimal policy from a set of expert demonstrations. While most IRL algorithms' theoretical guarantees rely on a linear reward structure, we aim to extend the theoretical understanding of IRL to scenarios where the reward function is parameterized by neural networks. Meanwhile, conventional IRL algorithms usually adopt a nested structure, leading to computational inefficiency, especially in high-dimensional settings. To address this problem, we propose the first two-timescale single-loop IRL algorithm under neural network parameterized reward and provide a non-asymptotic convergence analysis under overparameterization. Although prior optimality results for linear rewards do not apply, we show that our algorithm can identify the globally optimal reward and policy under certain neural network structures. This is the first IRL algorithm with a non-asymptotic convergence guarantee that provably achieves global optimality in neural network settings.

stat.ML↗

arXiv2025

From Demonstrations to Rewards: Alignment Without Explicit Human Preferences

One of the challenges of aligning large models with human preferences lies in both the data requirements and the technical complexities of current approaches. Predominant methods, such as RLHF, involve multiple steps, each demanding distinct types of data, including demonstration data and preference data. In RLHF, human preferences are typically modeled through a reward model, which serves as a proxy to guide policy learning during the reinforcement learning stage, ultimately producing a policy aligned with human preferences. However, in this paper, we propose a fresh perspective on learning alignment based on inverse reinforcement learning principles, where the optimal policy is still derived from reward maximization. However, instead of relying on preference data, we directly learn the reward model from demonstration data. This new formulation offers the flexibility to be applied even when only demonstration data is available, a capability that current RLHF methods lack, and it also shows that demonstration data offers more utility than what conventional wisdom suggests. Our extensive evaluation, based on public reward benchmark, HuggingFace Open LLM Leaderboard and MT-Bench, demonstrates that our approach compares favorably to state-of-the-art methods that rely solely on demonstration data.

cs.LG↗

arXiv2025

Bridging the Training-Inference Gap in LLMs by Leveraging Self-Generated Tokens

Language models are often trained to maximize the likelihood of the next token given past tokens in the training dataset. However, during inference time, they are utilized differently, generating text sequentially and auto-regressively by using previously generated tokens as input to predict the next one. Marginal differences in predictions at each step can cascade over successive steps, resulting in different distributions from what the models were trained for and potentially leading to unpredictable behavior. This paper proposes two simple approaches based on model own generation to address this discrepancy between the training and inference time. Our first approach is Batch-Scheduled Sampling, where, during training, we stochastically choose between the ground-truth token from the dataset and the model's own generated token as input to predict the next token. This is done in an offline manner, modifying the context window by interleaving ground-truth tokens with those generated by the model. Our second approach is Reference-Answer-based Correction, where we explicitly incorporate a self-correction capability into the model during training. This enables the model to effectively self-correct the gaps between the generated sequences and the ground truth data without relying on an external oracle model. By incorporating our proposed strategies during training, we have observed an overall improvement in performance compared to baseline methods, as demonstrated by our extensive experiments using summarization, general question-answering, and math question-answering tasks.

cs.LG↗

arXiv2024

Learning Reward and Policy Jointly from Demonstration and Preference Improves Alignment

Aligning human preference and value is an important requirement for building contemporary foundation models and embodied AI. However, popular approaches such as reinforcement learning with human feedback (RLHF) break down the task into successive stages, such as supervised fine-tuning (SFT), reward modeling (RM), and reinforcement learning (RL), each performing one specific learning task. Such a sequential approach results in serious issues such as significant under-utilization of data and distribution mismatch between the learned reward model and generated policy, which eventually lead to poor alignment performance. We develop a single stage approach named Alignment with Integrated Human Feedback (AIHF), capable of integrating both human preference and demonstration to train reward models and the policy. The proposed approach admits a suite of efficient algorithms, which can easily reduce to, and leverage, popular alignment algorithms such as RLHF and Directly Policy Optimization (DPO), and only requires minor changes to the existing alignment pipelines. We demonstrate the efficiency of the proposed solutions with extensive experiments involving alignment problems in LLMs and robotic control problems in MuJoCo. We observe that the proposed solutions outperform the existing alignment algorithms such as RLHF and DPO by large margins, especially when the amount of high-quality preference data is relatively limited.

cs.AI↗

arXiv2024

Getting More Juice Out of the SFT Data: Reward Learning from Human Demonstration Improves SFT for LLM Alignment

Aligning human preference and value is an important requirement for contemporary foundation models. State-of-the-art techniques such as Reinforcement Learning from Human Feedback (RLHF) often consist of two stages: 1) supervised fine-tuning (SFT), where the model is fine-tuned by learning from human demonstration data; 2) Preference learning, where preference data is used to learn a reward model, which is in turn used by a reinforcement learning (RL) step to fine-tune the model. Such reward model serves as a proxy to human preference, and it is critical to guide the RL step towards improving the model quality. In this work, we argue that the SFT stage significantly benefits from learning a reward model as well. Instead of using the human demonstration data directly via supervised learning, we propose to leverage an Inverse Reinforcement Learning (IRL) technique to simultaneously build an reward model and a policy model. This approach leads to new SFT algorithms that are not only efficient to implement, but are robust to the presence of low-quality supervised learning data. Moreover, we discover a connection between the proposed IRL based approach, and a recent line of works called Self-Play Fine-tune (SPIN). Theoretically, we show that the proposed algorithms converge to the stationary solutions of the IRL problem. Empirically, we align 1B and 7B models using proposed methods and evaluate them on a reward benchmark model and the HuggingFace Open LLM Leaderboard. The proposed methods show significant performance improvement over existing SFT approaches. Our results indicate that it is beneficial to leverage reward learning throughout the entire alignment process.

cs.AI↗

arXiv2024

A Bayesian Approach to Robust Inverse Reinforcement Learning

We consider a Bayesian approach to offline model-based inverse reinforcement learning (IRL). The proposed framework differs from existing offline model-based IRL approaches by performing simultaneous estimation of the expert's reward function and subjective model of environment dynamics. We make use of a class of prior distributions which parameterizes how accurate the expert's model of the environment is to develop efficient algorithms to estimate the expert's reward and subjective dynamics in high-dimensional settings. Our analysis reveals a novel insight that the estimated policy exhibits robust performance when the expert is believed (a priori) to have a highly accurate model of the environment. We verify this observation in the MuJoCo environments and show that our algorithms outperform state-of-the-art offline IRL algorithms.

cs.LG↗

arXiv2024

Structural Estimation of Markov Decision Processes in High-Dimensional State Space with Finite-Time Guarantees

We consider the task of estimating a structural model of dynamic decisions by a human agent based upon the observable history of implemented actions and visited states. This problem has an inherent nested structure: in the inner problem, an optimal policy for a given reward function is identified while in the outer problem, a measure of fit is maximized. Several approaches have been proposed to alleviate the computational burden of this nested-loop structure, but these methods still suffer from high complexity when the state space is either discrete with large cardinality or continuous in high dimensions. Other approaches in the inverse reinforcement learning (IRL) literature emphasize policy estimation at the expense of reduced reward estimation accuracy. In this paper we propose a single-loop estimation algorithm with finite time guarantees that is equipped to deal with high-dimensional state spaces without compromising reward estimation accuracy. In the proposed algorithm, each policy improvement step is followed by a stochastic gradient step for likelihood maximization. We show that the proposed algorithm converges to a stationary solution with a finite-time guarantee. Further, if the reward is parameterized linearly, we show that the algorithm approximates the maximum likelihood estimator sublinearly. Finally, by using robotics control problems in MuJoCo and their transfer settings, we show that the proposed algorithm achieves superior performance compared with other IRL and imitation learning benchmarks.

cs.LG↗

arXiv2024

When Demonstrations Meet Generative World Models: A Maximum Likelihood Framework for Offline Inverse Reinforcement Learning

Offline inverse reinforcement learning (Offline IRL) aims to recover the structure of rewards and environment dynamics that underlie observed actions in a fixed, finite set of demonstrations from an expert agent. Accurate models of expertise in executing a task has applications in safety-sensitive applications such as clinical decision making and autonomous driving. However, the structure of an expert's preferences implicit in observed actions is closely linked to the expert's model of the environment dynamics (i.e. the ``world'' model). Thus, inaccurate models of the world obtained from finite data with limited coverage could compound inaccuracy in estimated rewards. To address this issue, we propose a bi-level optimization formulation of the estimation task wherein the upper level is likelihood maximization based upon a conservative model of the expert's policy (lower level). The policy model is conservative in that it maximizes reward subject to a penalty that is increasing in the uncertainty of the estimated model of the world. We propose a new algorithmic framework to solve the bi-level optimization problem formulation and provide statistical and computational guarantees of performance for the associated optimal reward estimator. Finally, we demonstrate that the proposed algorithm outperforms the state-of-the-art offline IRL and imitation learning benchmarks by a large margin, over the continuous control tasks in MuJoCo and different datasets in the D4RL benchmark.

cs.LG↗

arXiv2022

Maximum-Likelihood Inverse Reinforcement Learning with Finite-Time Guarantees

Inverse reinforcement learning (IRL) aims to recover the reward function and the associated optimal policy that best fits observed sequences of states and actions implemented by an expert. Many algorithms for IRL have an inherently nested structure: the inner loop finds the optimal policy given parametrized rewards while the outer loop updates the estimates towards optimizing a measure of fit. For high dimensional environments such nested-loop structure entails a significant computational burden. To reduce the computational burden of a nested loop, novel methods such as SQIL [1] and IQ-Learn [2] emphasize policy estimation at the expense of reward estimation accuracy. However, without accurate estimated rewards, it is not possible to do counterfactual analysis such as predicting the optimal policy under different environment dynamics and/or learning new tasks. In this paper we develop a novel single-loop algorithm for IRL that does not compromise reward estimation accuracy. In the proposed algorithm, each policy improvement step is followed by a stochastic gradient step for likelihood maximization. We show that the proposed algorithm provably converges to a stationary solution with a finite-time guarantee. If the reward is parameterized linearly, we show the identified solution corresponds to the solution of the maximum entropy IRL problem. Finally, by using robotics control problems in MuJoCo and their transfer settings, we show that the proposed algorithm achieves superior performance compared with other IRL and imitation learning benchmarks.

cs.LG↗

arXiv2021

Learning to Coordinate in Multi-Agent Systems: A Coordinated Actor-Critic Algorithm and Finite-Time Guarantees

Multi-agent reinforcement learning (MARL) has attracted much research attention recently. However, unlike its single-agent counterpart, many theoretical and algorithmic aspects of MARL have not been well-understood. In this paper, we study the emergence of coordinated behavior by autonomous agents using an actor-critic (AC) algorithm. Specifically, we propose and analyze a class of coordinated actor-critic algorithms (CAC) in which individually parametrized policies have a {\it shared} part (which is jointly optimized among all agents) and a {\it personalized} part (which is only locally optimized). Such kind of {\it partially personalized} policy allows agents to learn to coordinate by leveraging peers' past experience and adapt to individual tasks. The flexibility in our design allows the proposed MARL-CAC algorithm to be used in a {\it fully decentralized} setting, where the agents can only communicate with their neighbors, as well as a {\it federated} setting, where the agents occasionally communicate with a server while optimizing their (partially personalized) local models. Theoretically, we show that under some standard regularity assumptions, the proposed MARL-CAC algorithm requires $\mathcal{O}(ε^{-\frac{5}{2}})$ samples to achieve an $ε$-stationary solution (defined as the solution whose squared norm of the gradient of the objective function is less than $ε$). To the best of our knowledge, this work provides the first finite-sample guarantee for decentralized AC algorithm with partially personalized policies.

cs.LG↗

arXiv2021

A Near-Optimal Algorithm for Stochastic Bilevel Optimization via Double-Momentum

This paper proposes a new algorithm -- the \underline{S}ingle-timescale Do\underline{u}ble-momentum \underline{St}ochastic \underline{A}pprox\underline{i}matio\underline{n} (SUSTAIN) -- for tackling stochastic unconstrained bilevel optimization problems. We focus on bilevel problems where the lower level subproblem is strongly-convex and the upper level objective function is smooth. Unlike prior works which rely on \emph{two-timescale} or \emph{double loop} techniques, we design a stochastic momentum-assisted gradient estimator for both the upper and lower level updates. The latter allows us to control the error in the stochastic gradient updates due to inaccurate solution to both subproblems. If the upper objective function is smooth but possibly non-convex, we show that {\aname}~requires $\mathcal{O}(ε^{-3/2})$ iterations (each using ${\cal O}(1)$ samples) to find an $ε$-stationary solution. The $ε$-stationary solution is defined as the point whose squared norm of the gradient of the outer function is less than or equal to $ε$. The total number of stochastic gradient samples required for the upper and lower level objective functions matches the best-known complexity for single-level stochastic gradient algorithms. We also analyze the case when the upper level objective function is strongly-convex.

math.OC↗

arXiv2020

On the Divergence of Decentralized Non-Convex Optimization

We study a generic class of decentralized algorithms in which $N$ agents jointly optimize the non-convex objective $f(u):=1/N\sum_{i=1}^{N}f_i(u)$, while only communicating with their neighbors. This class of problems has become popular in modeling many signal processing and machine learning applications, and many efficient algorithms have been proposed. However, by constructing some counter-examples, we show that when certain local Lipschitz conditions (LLC) on the local function gradient $\nabla f_i$'s are not satisfied, most of the existing decentralized algorithms diverge, even if the global Lipschitz condition (GLC) is satisfied, where the sum function $f$ has Lipschitz gradient. This observation raises an important open question: How to design decentralized algorithms when the LLC, or even the GLC, is not satisfied? To address the above question, we design a first-order algorithm called Multi-stage gradient tracking algorithm (MAGENTA), which is capable of computing stationary solutions with neither the LLC nor the GLC. In particular, we show that the proposed algorithm converges sublinearly to certain $ε$-stationary solution, where the precise rate depends on various algorithmic and problem parameters. In particular, if the local function $f_i$'s are $Q$th order polynomials, then the rate becomes $\mathcal{O}(1/ε^{Q-1})$. Such a rate is tight for the special case of $Q=2$ where each $f_i$ satisfies LLC. To our knowledge, this is the first attempt that studies decentralized non-convex optimization problems with neither the LLC nor the GLC.

math.OC↗