Lihong Li

Zhepei Wei, Wenlin Yao, Yao Liu et al.

While reinforcement learning (RL) has demonstrated remarkable success in enhancing large language models (LLMs), it has primarily focused on single-turn tasks such as solving math problems. Training effective web agents for multi-turn interactions remains challenging due to the complexity of long-horizon decision-making across dynamic web interfaces. In this work, we present WebAgent-R1, a simple yet effective end-to-end multi-turn RL framework for training web agents. It learns directly from online interactions with web environments by asynchronously generating diverse trajectories, entirely guided by binary rewards depending on task success. Experiments on the WebArena-Lite benchmark demonstrate the effectiveness of WebAgent-R1, boosting the task success rate of Qwen-2.5-3B from 6.1% to 33.9% and Llama-3.1-8B from 8.5% to 44.8%, significantly outperforming existing state-of-the-art methods and strong proprietary models such as OpenAI o3. In-depth analyses reveal the effectiveness of the thinking-based prompting strategy and test-time scaling through increased interactions for web tasks. We further investigate different RL initialization policies by introducing two variants, namely WebAgent-R1-Zero and WebAgent-R1-CoT, which highlight the importance of the warm-up training stage (i.e., behavior cloning) and provide insights on incorporating long chain-of-thought (CoT) reasoning in web agents.

Jiacheng Lin, Zhongruo Wang, Kun Qian et al.

Supervised Fine-Tuning (SFT) on domain-specific datasets is a common approach to adapt Large Language Models (LLMs) to specialized tasks but is often believed to degrade their general capabilities. In this work, we revisit this trade-off and present both empirical and theoretical insights. First, we show that SFT does not always hurt: using a smaller learning rate can substantially mitigate general performance degradation while preserving comparable target-domain performance. We then provide a theoretical analysis that explains these phenomena and further motivates a new method, Token-Adaptive Loss Reweighting (TALR). Building on this, and recognizing that smaller learning rates alone do not fully eliminate general-performance degradation in all cases, we evaluate a range of strategies for reducing general capability loss, including L2 regularization, LoRA, model averaging, FLOW, and our proposed TALR. Experimental results demonstrate that while no method completely eliminates the trade-off, TALR consistently outperforms these baselines in balancing domain-specific gains and general capabilities. Finally, we distill our findings into practical guidelines for adapting LLMs to new domains: (i) using a small learning rate to achieve a favorable trade-off, and (ii) when a stronger balance is further desired, adopt TALR as an effective strategy.

Xiaoyu Chen, Jiachen Hu, Chi Jin et al.

Reinforcement learning encounters many challenges when applied directly in the real world. Sim-to-real transfer is widely used to transfer the knowledge learned from simulation to the real world. Domain randomization -- one of the most popular algorithms for sim-to-real transfer -- has been demonstrated to be effective in various tasks in robotics and autonomous driving. Despite its empirical successes, theoretical understanding on why this simple algorithm works is limited. In this paper, we propose a theoretical framework for sim-to-real transfers, in which the simulator is modeled as a set of MDPs with tunable parameters (corresponding to unknown physical parameters such as friction). We provide sharp bounds on the sim-to-real gap -- the difference between the value of policy returned by domain randomization and the value of an optimal policy for the real world. We prove that sim-to-real transfer can succeed under mild conditions without any real-world training samples. Our theory also highlights the importance of using memory (i.e., history-dependent policies) in domain randomization. Our proof is based on novel techniques that reduce the problem of bounding the sim-to-real gap to the problem of designing efficient learning algorithms for infinite-horizon MDPs, which we believe are of independent interest.

Yi Liu, Lihong Li

The rich body of Bandit literature not only offers a diverse toolbox of algorithms, but also makes it hard for a practitioner to find the right solution to solve the problem at hand. Typical textbooks on Bandits focus on designing and analyzing algorithms, and surveys on applications often present a list of individual applications. While these are valuable resources, there exists a gap in mapping applications to appropriate Bandit algorithms. In this paper, we aim to reduce this gap with a structured map of Bandits to help practitioners navigate to find relevant and practical Bandit algorithms. Instead of providing a comprehensive overview, we focus on a small number of key decision points related to reward, action, and features, which often affect how Bandit algorithms are chosen in practice.

Xiaoyu Chen, Jiachen Hu, Lihong Li et al.

Reinforcement learning (RL) in episodic, factored Markov decision processes (FMDPs) is studied. We propose an algorithm called FMDP-BF, which leverages the factorization structure of FMDP. The regret of FMDP-BF is shown to be exponentially smaller than that of optimal algorithms designed for non-factored MDPs, and improves on the best previous result for FMDPs~\citep{osband2014near} by a factored of $\sqrt{H|\mathcal{S}_i|}$, where $|\mathcal{S}_i|$ is the cardinality of the factored state subspace and $H$ is the planning horizon. To show the optimality of our bounds, we also provide a lower bound for FMDP, which indicates that our algorithm is near-optimal w.r.t. timestep $T$, horizon $H$ and factored state-action subspace cardinality. Finally, as an application, we study a new formulation of constrained RL, known as RL with knapsack constraints (RLwK), and provides the first sample-efficient algorithm based on FMDP-BF.

Ali Mousavi, Lihong Li, Qiang Liu et al.

Off-policy estimation for long-horizon problems is important in many real-life applications such as healthcare and robotics, where high-fidelity simulators may not be available and on-policy evaluation is expensive or impossible. Recently, \cite{liu18breaking} proposed an approach that avoids the \emph{curse of horizon} suffered by typical importance-sampling-based methods. While showing promising results, this approach is limited in practice as it requires data be drawn from the \emph{stationary distribution} of a \emph{known} behavior policy. In this work, we propose a novel approach that eliminates such limitations. In particular, we formulate the problem as solving for the fixed point of a certain operator. Using tools from Reproducing Kernel Hilbert Spaces (RKHSs), we develop a new estimator that computes importance ratios of stationary distributions, without knowledge of how the off-policy data are collected. We analyze its asymptotic consistency and finite-sample generalization. Experiments on benchmarks verify the effectiveness of our approach.

Ge Liu, Rui Wu, Heng-Tze Cheng et al.

Deep Reinforcement Learning (RL) is proven powerful for decision making in simulated environments. However, training deep RL model is challenging in real world applications such as production-scale health-care or recommender systems because of the expensiveness of interaction and limitation of budget at deployment. One aspect of the data inefficiency comes from the expensive hyper-parameter tuning when optimizing deep neural networks. We propose Adaptive Behavior Policy Sharing (ABPS), a data-efficient training algorithm that allows sharing of experience collected by behavior policy that is adaptively selected from a pool of agents trained with an ensemble of hyper-parameters. We further extend ABPS to evolve hyper-parameters during training by hybridizing ABPS with an adapted version of Population Based Training (ABPS-PBT). We conduct experiments with multiple Atari games with up to 16 hyper-parameter/architecture setups. ABPS achieves superior overall performance, reduced variance on top 25% agents, and equivalent performance on the best agent compared to conventional hyper-parameter tuning with independent training, even though ABPS only requires the same number of environmental interactions as training a single agent. We also show that ABPS-PBT further improves the convergence speed and reduces the variance.

Christoph Dann, Lihong Li, Wei Wei et al.

The performance of a reinforcement learning algorithm can vary drastically during learning because of exploration. Existing algorithms provide little information about the quality of their current policy before executing it, and thus have limited use in high-stakes applications like healthcare. We address this lack of accountability by proposing that algorithms output policy certificates. These certificates bound the sub-optimality and return of the policy in the next episode, allowing humans to intervene when the certified quality is not satisfactory. We further introduce two new algorithms with certificates and present a new framework for theoretical analysis that guarantees the quality of their policies and certificates. For tabular MDPs, we show that computing certificates can even improve the sample-efficiency of optimism-based exploration. As a result, one of our algorithms is the first to achieve minimax-optimal PAC bounds up to lower-order terms, and this algorithm also matches (and in some settings slightly improves upon) existing minimax regret bounds.

Kwang-Sung Jun, Lihong Li, Yuzhe Ma et al.

We study adversarial attacks that manipulate the reward signals to control the actions chosen by a stochastic multi-armed bandit algorithm. We propose the first attack against two popular bandit algorithms: $ε$-greedy and UCB, \emph{without} knowledge of the mean rewards. The attacker is able to spend only logarithmic effort, multiplied by a problem-specific parameter that becomes smaller as the bandit problem gets easier to attack. The result means the attacker can easily hijack the behavior of the bandit algorithm to promote or obstruct certain actions, say, a particular medical treatment. As bandits are seeing increasingly wide use in practice, our study exposes a significant security threat.

Yuzhe Ma, Kwang-Sung Jun, Lihong Li et al.

We study offline data poisoning attacks in contextual bandits, a class of reinforcement learning problems with important applications in online recommendation and adaptive medical treatment, among others. We provide a general attack framework based on convex optimization and show that by slightly manipulating rewards in the data, an attacker can force the bandit algorithm to pull a target arm for a target contextual vector. The target arm and target contextual vector are both chosen by the attacker. That is, the attacker can hijack the behavior of a contextual bandit. We also investigate the feasibility and the side effects of such attacks, and identify future directions for defense. Experiments on both synthetic and real-world data demonstrate the efficiency of the attack algorithm.

Yichen Chen, Lihong Li, Mengdi Wang

Approximate linear programming (ALP) represents one of the major algorithmic families to solve large-scale Markov decision processes (MDP). In this work, we study a primal-dual formulation of the ALP, and develop a scalable, model-free algorithm called bilinear $π$ learning for reinforcement learning when a sampling oracle is provided. This algorithm enjoys a number of advantages. First, it adopts (bi)linear models to represent the high-dimensional value function and state-action distributions, using given state and action features. Its run-time complexity depends on the number of features, not the size of the underlying MDPs. Second, it operates in a fully online fashion without having to store any sample, thus having minimal memory footprint. Third, we prove that it is sample-efficient, solving for the optimal policy to high precision with a sample complexity linear in the dimension of the parameter space.

Baolin Peng, Xiujun Li, Lihong Li et al.

Building a dialogue agent to fulfill complex tasks, such as travel planning, is challenging because the agent has to learn to collectively complete multiple subtasks. For example, the agent needs to reserve a hotel and book a flight so that there leaves enough time for commute between arrival and hotel check-in. This paper addresses this challenge by formulating the task in the mathematical framework of options over Markov Decision Processes (MDPs), and proposing a hierarchical deep reinforcement learning approach to learning a dialogue manager that operates at different temporal scales. The dialogue manager consists of: (1) a top-level dialogue policy that selects among subtasks or options, (2) a low-level dialogue policy that selects primitive actions to complete the subtask given by the top-level policy, and (3) a global state tracker that helps ensure all cross-subtask constraints be satisfied. Experiments on a travel planning task with simulated and real users show that our approach leads to significant improvements over three baselines, two based on handcrafted rules and the other based on flat deep reinforcement learning.

Lihong Li, Yu Lu, Dengyong Zhou

Contextual bandits are widely used in Internet services from news recommendation to advertising, and to Web search. Generalized linear models (logistical regression in particular) have demonstrated stronger performance than linear models in many applications where rewards are binary. However, most theoretical analyses on contextual bandits so far are on linear bandits. In this work, we propose an upper confidence bound based algorithm for generalized linear contextual bandits, which achieves an $\tilde{O}(\sqrt{dT})$ regret over $T$ rounds with $d$ dimensional feature vectors. This regret matches the minimax lower bound, up to logarithmic terms, and improves on the best previous result by a $\sqrt{d}$ factor, assuming the number of arms is fixed. A key component in our analysis is to establish a new, sharp finite-sample confidence bound for maximum-likelihood estimates in generalized linear models, which may be of independent interest. We also analyze a simpler upper confidence bound algorithm, which is useful in practice, and prove it to have optimal regret for certain cases.

Xiujun Li, Zachary C. Lipton, Bhuwan Dhingra et al.

Despite widespread interests in reinforcement-learning for task-oriented dialogue systems, several obstacles can frustrate research and development progress. First, reinforcement learners typically require interaction with the environment, so conventional dialogue corpora cannot be used directly. Second, each task presents specific challenges, requiring separate corpus of task-specific annotated data. Third, collecting and annotating human-machine or human-human conversations for task-oriented dialogues requires extensive domain knowledge. Because building an appropriate dataset can be both financially costly and time-consuming, one popular approach is to build a user simulator based upon a corpus of example dialogues. Then, one can train reinforcement learning agents in an online fashion as they interact with the simulator. Dialogue agents trained on these simulators can serve as an effective starting point. Once agents master the simulator, they may be deployed in a real environment to interact with humans, and continue to be trained online. To ease empirical algorithmic comparisons in dialogues, this paper introduces a new, publicly available simulation framework, where our simulator, designed for the movie-booking domain, leverages both rules and collected data. The simulator supports two tasks: movie ticket booking and movie seeking. Finally, we demonstrate several agents and detail the procedure to add and test your own agent in the proposed framework.

Zachary C. Lipton, Kamyar Azizzadenesheli, Abhishek Kumar et al.

Many practical environments contain catastrophic states that an optimal agent would visit infrequently or never. Even on toy problems, Deep Reinforcement Learning (DRL) agents tend to periodically revisit these states upon forgetting their existence under a new policy. We introduce intrinsic fear (IF), a learned reward shaping that guards DRL agents against periodic catastrophes. IF agents possess a fear model trained to predict the probability of imminent catastrophe. This score is then used to penalize the Q-learning objective. Our theoretical analysis bounds the reduction in average return due to learning on the perturbed objective. We also prove robustness to classification errors. As a bonus, IF models tend to learn faster, owing to reward shaping. Experiments demonstrate that intrinsic-fear DQNs solve otherwise pathological environments and improve on several Atari games.

Emma Brunskill, Lihong Li

Transferring knowledge across a sequence of related tasks is an important challenge in reinforcement learning (RL). Despite much encouraging empirical evidence, there has been little theoretical analysis. In this paper, we study a class of lifelong RL problems: the agent solves a sequence of tasks modeled as finite Markov decision processes (MDPs), each of which is from a finite set of MDPs with the same state/action sets and different transition/reward functions. Motivated by the need for cross-task exploration in lifelong learning, we formulate a novel online coupon-collector problem and give an optimal algorithm. This allows us to develop a new lifelong RL algorithm, whose overall sample complexity in a sequence of tasks is much smaller than single-task learning, even if the sequence of tasks is generated by an adversary. Benefits of the algorithm are demonstrated in simulated problems, including a recently introduced human-robot interaction problem.

Shipra Agrawal, Nikhil R. Devanur, Lihong Li

We consider a contextual version of multi-armed bandit problem with global knapsack constraints. In each round, the outcome of pulling an arm is a scalar reward and a resource consumption vector, both dependent on the context, and the global knapsack constraints require the total consumption for each resource to be below some pre-fixed budget. The learning agent competes with an arbitrary set of context-dependent policies. This problem was introduced by Badanidiyuru et al. (2014), who gave a computationally inefficient algorithm with near-optimal regret bounds for it. We give a computationally efficient algorithm for this problem with slightly better regret bounds, by generalizing the approach of Agarwal et al. (2014) for the non-constrained version of the problem. The computational time of our algorithm scales logarithmically in the size of the policy space. This answers the main open question of Badanidiyuru et al. (2014). We also extend our results to a variant where there are no knapsack constraints but the objective is an arbitrary Lipschitz concave function of the sum of outcome vectors.

Dragomir Yankov, Pavel Berkhin, Lihong Li

We analyze the problem of using Explore-Exploit techniques to improve precision in multi-result ranking systems such as web search, query autocompletion and news recommendation. Adopting an exploration policy directly online, without understanding its impact on the production system, may have unwanted consequences - the system may sustain large losses, create user dissatisfaction, or collect exploration data which does not help improve ranking quality. An offline framework is thus necessary to let us decide what policy and how we should apply in a production environment to ensure positive outcome. Here, we describe such an offline framework. Using the framework, we study a popular exploration policy - Thompson sampling. We show that there are different ways of implementing it in multi-result ranking systems, each having different semantic interpretation and leading to different results in terms of sustained click-through-rate (CTR) loss and expected model improvement. In particular, we demonstrate that Thompson sampling can act as an online learner optimizing CTR, which in some cases can lead to an interesting outcome: lift in CTR during exploration. The observation is important for production systems as it suggests that one can get both valuable exploration data to improve ranking performance on the long run, and at the same time increase CTR while exploration lasts.

Lihong Li, Remi Munos, Csaba Szepesvari

This paper studies the off-policy evaluation problem, where one aims to estimate the value of a target policy based on a sample of observations collected by another policy. We first consider the multi-armed bandit case, establish a minimax risk lower bound, and analyze the risk of two standard estimators. It is shown, and verified in simulation, that one is minimax optimal up to a constant, while another can be arbitrarily worse, despite its empirical success and popularity. The results are applied to related problems in contextual bandits and fixed-horizon Markov decision processes, and are also related to semi-supervised learning.

Lihong Li, Shunbao Chen, Jim Kleban et al.

Optimizing an interactive system against a predefined online metric is particularly challenging, when the metric is computed from user feedback such as clicks and payments. The key challenge is the counterfactual nature: in the case of Web search, any change to a component of the search engine may result in a different search result page for the same query, but we normally cannot infer reliably from search log how users would react to the new result page. Consequently, it appears impossible to accurately estimate online metrics that depend on user feedback, unless the new engine is run to serve users and compared with a baseline in an A/B test. This approach, while valid and successful, is unfortunately expensive and time-consuming. In this paper, we propose to address this problem using causal inference techniques, under the contextual-bandit framework. This approach effectively allows one to run (potentially infinitely) many A/B tests offline from search log, making it possible to estimate and optimize online metrics quickly and inexpensively. Focusing on an important component in a commercial search engine, we show how these ideas can be instantiated and applied, and obtain very promising results that suggest the wide applicability of these techniques.

Zhen Qin, Vaclav Petricek, Nikos Karampatziakis et al.

Bootstrapping is a useful technique for estimating the uncertainty of a predictor, for example, confidence intervals for prediction. It is typically used on small to moderate sized datasets, due to its high computation cost. This work describes a highly scalable online bootstrapping strategy, implemented inside Vowpal Wabbit, that is several times faster than traditional strategies. Our experiments indicate that, in addition to providing a black box-like method for estimating uncertainty, our implementation of online bootstrapping may also help to train models with better prediction performance due to model averaging.

Lihong Li

Thompson Sampling, one of the oldest heuristics for solving multi-armed bandits, has recently been shown to demonstrate state-of-the-art performance. The empirical success has led to great interests in theoretical understanding of this heuristic. In this paper, we approach this problem in a way very different from existing efforts. In particular, motivated by the connection between Thompson Sampling and exponentiated updates, we propose a new family of algorithms called Generalized Thompson Sampling in the expert-learning framework, which includes Thompson Sampling as a special case. Similar to most expert-learning algorithms, Generalized Thompson Sampling uses a loss function to adjust the experts' weights. General regret bounds are derived, which are also instantiated to two important loss functions: square loss and logarithmic loss. In contrast to existing bounds, our results apply to quite general contextual bandits. More importantly, they quantify the effect of the "prior" distribution on the regret bounds.

Miroslav Dudik, Dumitru Erhan, John Langford et al.

We present and prove properties of a new offline policy evaluator for an exploration learning setting which is superior to previous evaluators. In particular, it simultaneously and correctly incorporates techniques from importance weighting, doubly robust evaluation, and nonstationary policy evaluation approaches. In addition, our approach allows generating longer histories by careful control of a bias-variance tradeoff, and further decreases variance by incorporating information about randomness of the target policy. Empirical evidence from synthetic and realworld exploration learning problems shows the new evaluator successfully unifies previous approaches and uses information an order of magnitude more efficiently.

Alexander L. Strehl, Lihong Li, Michael L. Littman

Model-based learning algorithms have been shown to use experience efficiently when learning to solve Markov Decision Processes (MDPs) with finite state and action spaces. However, their high computational cost due to repeatedly solving an internal model inhibits their use in large-scale problems. We propose a method based on real-time dynamic programming (RTDP) to speed up two model-based algorithms, RMAX and MBIE (model-based interval estimation), resulting in computationally much faster algorithms with little loss compared to existing bounds. Specifically, our two new learning algorithms, RTDP-RMAX and RTDP-IE, have considerably smaller computational demands than RMAX and MBIE. We develop a general theoretical framework that allows us to prove that both are efficient learners in a PAC (probably approximately correct) sense. We also present an experimental evaluation of these new algorithms that helps quantify the tradeoff between computational and experience demands.

Emma Brunskill, Bethany Leffler, Lihong Li et al.

Continuous state spaces and stochastic, switching dynamics characterize a number of rich, realworld domains, such as robot navigation across varying terrain. We describe a reinforcementlearning algorithm for learning in these domains and prove for certain environments the algorithm is probably approximately correct with a sample complexity that scales polynomially with the state-space dimension. Unfortunately, no optimal planning techniques exist in general for such problems; instead we use fitted value iteration to solve the learned MDP, and include the error due to approximate planning in our bounds. Finally, we report an experiment using a robotic car driving over varying terrain to demonstrate that these dynamics representations adequately capture real-world dynamics and that our algorithm can be used to efficiently solve such problems.