It is often of interest to learn a context-sensitive decision policy, such as in contextual multi-armed bandit processes. To quantify the efficiency of a machine learning algorithm for such settings, probably approximately correct (PAC) bounds, which bound the number of samples required, or cumulative regret guarantees, are typically used. However, real-world settings often have limited resources for experimentation, and decisions/interventions may differ in the amount of resources required (e.g., money or time). Therefore, it is of interest to consider how to design an experiment strategy that reduces the experimental budget needed to learn a near-optimal contextual policy. Unlike reinforcement learning or bandit approaches that embed costs into the reward function, we focus on reducing resource use in learning a near-optimal policy without resource constraints. We introduce two resource-aware algorithms for the contextual bandit setting and prove their soundness. Simulations based on real-world datasets demonstrate that our algorithms significantly reduce the resources needed to learn a near-optimal decision policy compared to previous resource-unaware methods.

Thang Luong, Dawsen Hwang, Hoang H Nguyen, Golnaz Ghiasi, Yuri Chervonyi, Insuk Seo, Junsu Kim, Garrett Bingham, Jonathan Lee, Swaroop Mishra, Alex Zhai, Huiyi Hu, Henryk Michalewski, Jimin Kim, Jeonghyun Ahn, Junhwi Bae, Xingyou Song, Trieu Hoang Trinh, Quoc V Le, Junehyuk Jung. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Despite their success in many domains, large language models (LLMs) remain under-studied in scenarios requiring optimal decision-making under uncertainty. This is crucial as many real-world applications, ranging from personalized recommendations to healthcare interventions, demand that LLMs not only predict but also actively learn to make optimal decisions through exploration. In this work, we measure LLMs’ (in)ability to make optimal decisions in bandits, a state-less reinforcement learning setting relevant to many applications. We develop a comprehensive suite of environments, including both context-free and contextual bandits with varying task difficulties, to benchmark LLMs’ performance. Motivated by the existence of optimal exploration algorithms, we propose efficient ways to integrate this algorithmic knowledge into LLMs: by providing explicit algorithm-guided support during inference; and through algorithm distillation via in-context demonstrations and fine-tuning, using synthetic data generated from these algorithms. Impressively, these techniques allow us to achieve superior exploration performance with smaller models, surpassing larger models on various tasks. We conducted an extensive ablation study to shed light on various factors, such as task difficulty and data representation, that influence the efficiency of LLM exploration. Additionally, we conduct a rigorous analysis of the LLM’s exploration efficiency using the concept of regret, linking its ability to explore to the model size and underlying algorithm.

We consider the problem of preference-based reinforcement learning (PbRL), where, unlike traditional reinforcement learning (RL), an agent receives feedback only in terms of 1 bit (0/1) preferences over a trajectory pair instead of absolute rewards for it. The success of the traditional reward-based RL framework crucially depends on how accurately a system designer can express an appropriate reward function, which is often a non-trivial task. The main novelty of the our framework is the ability to learn from preference-based trajectory feedback that eliminates the need to hand-craft numeric reward models. This paper sets up a formal framework for the PbRL problem with non-markovian rewards, where the trajectory preferences are encoded by a generalized linear model of dimension $d$. Assuming the transition model is known, we propose an algorithm with a regret guarantee of $\tilde {\mathcal{O}}\left( SH d \log (T / \delta) \sqrt{T} \right)$. We further extend the above algorithm to the case of unknown transition dynamics and provide an algorithm with regret $\widetilde{\mathcal{O}}((\sqrt{d} + H^2 + |\mathcal{S}|)\sqrt{dT} +\sqrt{|\mathcal{S}||\mathcal{A}|TH} )$. To the best of our knowledge, our work is one of the first to give tight regret guarantees for preference-based RL problem with trajectory preferences.

POMDPs capture a broad class of decision making problems, but hardness results suggest that learning is intractable even in simple settings due to the inherent partial observability. However, in many realistic problems, more information is either revealed or can be computed during some point of the learning process. Motivated by diverse applications ranging from robotics to data center scheduling, we formulate a Hindsight Observable Markov Decision Process (HOMDP) as a POMDP where the latent states are revealed to the learner in hindsight and only during training. We introduce new algorithms for the tabular and function approximation settings that are provably sample-efficient with hindsight observability, even in POMDPs that would otherwise be statistically intractable. We give a lower bound showing that the tabular algorithm is optimal in its dependence on latent state and observation cardinalities.

Large transformer models trained on diverse datasets have shown a remarkable ability to learn in-context, achieving high few-shot performance on tasks they were not explicitly trained to solve. In this paper, we study the in-context learning capabilities of transformers in decision-making problems, i.e., reinforcement learning (RL) for bandits and Markov decision processes. To do so, we introduce and study the Decision-Pretrained Transformer (DPT), a supervised pretraining method where a transformer predicts an optimal action given a query state and an in-context dataset of interactions from a diverse set of tasks. While simple, this procedure produces a model with several surprising capabilities. We find that the trained transformer can solve a range of RL problems in-context, exhibiting both exploration online and conservatism offline, despite not being explicitly trained to do so. The model also generalizes beyond the pretraining distribution to new tasks and automatically adapts its decision-making strategies to unknown structure. Theoretically, we show DPT can be viewed as an efficient implementation of Bayesian posterior sampling, a provably sample-efficient RL algorithm. We further leverage this connection to provide guarantees on the regret of the in-context algorithm yielded by DPT, and prove that it can learn faster than algorithms used to generate the pretraining data. These results suggest a promising yet simple path towards instilling strong in-context decision-making abilities in transformers.

We study the problem of experiment planning with function approximation in contextual bandit problems. In settings where there is a significant overhead to deploying adaptive algorithms---for example, when the execution of the data collection policies is required to be distributed, or a human in the loop is needed to implement these policies---producing in advance a set of policies for data collection is paramount. We study the setting where a large dataset of contexts but not rewards is available and may be used by the learner to design an effective data collection strategy. Although when rewards are linear this problem has been well studied, results are still missing for more complex reward models. In this work we propose two experiment planning strategies compatible with function approximation. The first is an eluder planning and sampling procedure that can recover optimality guarantees depending on the eluder dimension of the reward function class. For the second, we show that a uniform sampler achieves competitive optimality rates in the setting where the number of actions is small. We finalize our results introducing a statistical gap fleshing out the fundamental differences between planning and adaptive learning and provide results for planning with model selection.

We study the problem of model selection in batch policy optimization: given a fixed, partial-feedback dataset and M model classes, learn a policy with performance that is competitive with the policy derived from the best model class. We formalize the problem in the contextual bandit setting with linear model classes by identifying three sources of error that any model selection algorithm should optimally trade-off in order to be competitive: (1) approximation error, (2) statistical complexity, and (3) coverage. The first two sources are common in model selection for supervised learning, where optimally trading off these two is well-studied. In contrast, the third source is unique to batch policy optimization and is due to dataset shift inherent to the setting. We first show that no batch policy optimization algorithm can achieve a guarantee addressing all three simultaneously, revealing a stark contrast between difficulties in batch policy optimization and the positive results available in supervised learning. Despite this negative result, we show that relaxing any one of the three error sources enables the design of algorithms achieving near-oracle inequalities for the remaining two. We conclude with experiments demonstrating the efficacy of these algorithms.

In offline reinforcement learning (RL), a learner leverages prior logged data to learn a good policy without interacting with the environment. A major challenge in applying such methods in practice is the lack of both theoretically principled and practical tools for model selection and evaluation. To address this, we study the problem of model selection in offline RL with value function approximation. The learner is given a nested sequence of model classes to minimize squared Bellman error and must select among these to achieve a balance between approximation and estimation error of the classes. We propose the first model selection algorithm for offline RL that achieves minimax rate-optimal oracle inequalities up to logarithmic factors. The algorithm, ModBE, takes as input a collection of candidate model classes and a generic base offline RL algorithm. By successively eliminating model classes using a novel one-sided generalization test, ModBE returns a policy with regret scaling with the complexity of the minimally complete model class. In addition to its theoretical guarantees, it is conceptually simple and computationally efficient, amounting to solving a series of square loss regression problems and then comparing relative square loss between classes. We conclude with several numerical simulations showing it is capable of reliably selecting a good model class.

Deep reinforcement learning has achieved impressive successes yet often requires a very large amount of interaction data. This result is perhaps unsurprising, as using complicated function approximation often requires more data to fit, and early theoretical results on linear Markov decision processes provide regret bounds that scale with the dimension of the linear approximation. Ideally, we would like to automatically identify the minimal dimension of the approximation that is sufficient to encode an optimal policy. Towards this end, we consider the problem of model selection in RL with function approximation, given a set of candidate RL algorithms with known regret guarantees. The learner’s goal is to adapt to the complexity of the optimal algorithm without knowing it a priori. We present a meta-algorithm that successively rejects increasingly complex models using a simple statistical test. Given at least one candidate that satisfies realizability, we prove the meta-algorithm adapts to the optimal complexity with regret that is only marginally suboptimal in the number of episodes and number of candidate algorithms. The dimension and horizon dependencies remain optimal with respect to the best candidate, and our meta-algorithmic approach is flexible to incorporate multiple candidate algorithms and models. Finally, we show that the meta-algorithm automatically admits significantly improved instance-dependent regret bounds that depend on the gaps between the maximal values attainable by the candidates.

Since its introduction a decade ago, relative entropy policy search (REPS) has demonstrated successful policy learning on a number of simulated and real-world robotic domains, not to mention providing algorithmic components used by many recently proposed reinforcement learning (RL) algorithms. While REPS is commonly known in the community, there exist no guarantees on its performance when using stochastic and gradient-based solvers. In this paper we aim to fill this gap by providing guarantees and convergence rates for the sub-optimality of a policy learned using first-order optimization methods applied to the REPS objective. We first consider the setting in which we are given access to exact gradients and demonstrate how near-optimality of the objective translates to near-optimality of the policy. We then consider the practical setting of stochastic gradients, and introduce a technique that uses generative access to the underlying Markov decision process to compute parameter updates that maintain favorable convergence to the optimal regularized policy.

In the stochastic linear contextual bandit setting there exist several minimax procedures for exploration with policies that are reactive to the data being acquired. In practice, there can be a significant engineering overhead to deploy these algorithms, especially when the dataset is collected in a distributed fashion or when a human in the loop is needed to implement a different policy. Exploring with a single non-reactive policy is beneficial in such cases. Assuming some batch contexts are available, we design a single stochastic policy to collect a good dataset from which a near-optimal policy can be extracted. We present a theoretical analysis as well as numerical experiments on both synthetic and real-world datasets.

Online learning is a powerful tool for analyzing iterative algorithms. However, the classic adversarial setup fails to capture regularity that can exist in practice. Motivated by this observation, we establish a new setup, called Continuous Online Learning (COL), where the gradient of online loss function changes continuously across rounds with respect to the learner’s decisions. We show that COL appropriately describes many interesting applications, from general equilibrium problems (EPs) to optimization in episodic MDPs. Using this new setup, we revisit the difficulty of sublinear dynamic regret. We prove a fundamental equivalence between achieving sublinear dynamic regret in COL and solving certain EPs. With this insight, we offer conditions for efficient algorithms that achieve sublinear dynamic regret, even when the losses are chosen adaptively without any a priori variation budget. Furthermore, we show for COL a reduction from dynamic regret to both static regret and convergence in the associated EP, allowing us to analyze the dynamic regret of many existing algorithms.

Maximum a posteriori (MAP) inference is a fundamental computational paradigm for statistical inference. In the setting of graphical models, MAP inference entails solving a combinatorial optimization problem to find the most likely configuration of the discrete-valued model. Linear programming (LP) relaxations in the Sherali-Adams hierarchy are widely used to attempt to solve this problem, and smooth message passing algorithms have been proposed to solve regularized versions of these LPs with great success. This paper leverages recent work in entropy-regularized LPs to analyze convergence rates of a class of edge-based smooth message passing algorithms to epsilon-optimality in the relaxation. With an appropriately chosen regularization constant, we present a theoretical guarantee on the number of iterations sufficient to recover the true integral MAP solution when the LP is tight and the solution is unique.

Maximum a posteriori (MAP) inference in discrete-valued Markov random fields is a fundamental problem in machine learning that involves identifying the most likely configuration of random variables given a distribution. Due to the difficulty of this combinatorial problem, linear programming (LP) relaxations are commonly used to derive specialized message passing algorithms that are often interpreted as coordinate descent on the dual LP. To achieve more desirable computational properties, a number of methods regularize the LP with an entropy term, leading to a class of smooth message passing algorithms with convergence guarantees. In this paper, we present randomized methods for accelerating these algorithms by leveraging techniques that underlie classical accelerated gradient methods. The proposed algorithms incorporate the familiar steps of standard smooth message passing algorithms, which can be viewed as coordinate minimization steps. We show that these accelerated variants achieve faster rates for finding $\epsilon$-optimal points of the unregularized problem, and, when the LP is tight, we prove that the proposed algorithms recover the true MAP solution in fewer iterations than standard message passing algorithms.

On-policy imitation learning algorithms such as Dagger evolve a robot control policy by executing it, measuring performance (loss), obtaining corrective feedback from a supervisor, and generating the next policy. As the loss between iterations can vary unpredictably, a fundamental question is under what conditions this process will eventually achieve a converged policy. If one assumes the underlying trajectory distribution is static (stationary), it is possible to prove convergence for Dagger. Cheng and Boots (2018) consider the more realistic model for robotics where the underlying trajectory distribution, which is a function of the policy, is dynamic and show that it is possible to prove convergence when a condition on the rate of change of the trajectory distributions is satisfied. In this paper, we reframe that result using dynamic regret theory from the field of Online Optimization to prove convergence to locally optimal policies for Dagger, Imitation Gradient, and Multiple Imitation Gradient. These results inspire a new algorithm, Adaptive On-Policy Regularization (AOR), that ensures the conditions for convergence. We present simulation results with cart-pole balancing and walker locomotion benchmarks that suggest AOR can significantly decrease dynamic regret and chattering. To our knowledge, this the first application of dynamic regret theory to imitation learning.

Generalizing manipulation skills to new situations requires extracting invariant patterns from demonstrations. For example, the robot needs to understand the demonstrations at a higher level while being invariant to the appearance of the objects, geometric aspects of objects such as its position, size, orientation and viewpoint of the observer in the demonstrations. In this paper, we propose an algorithm that learns a joint probability density function of the demonstrations with invariant formulations of hidden semi-Markov models to extract invariant segments (also called sub-goals or options), and smoothly follow the generated sequence of states with a linear quadratic tracking controller. The algorithm takes as input the demonstrations observed with respect to different coordinate systems describing virtual landmarks or objects of interest, and adapts the segments according to the environmental changes in a systematic manner. We present variants of this algorithm in latent space with low-rank covariance decompositions, semi-tied covariances, and non-parametric online estimation of model parameters under small variance asymptotics; yielding considerably low sample and model complexity for acquiring new manipulation skills. The algorithm allows a Baxter robot to learn a pick-and-place task while avoiding a movable obstacle based on only 4 kinesthetic demonstrations.

Motivated by recent advances in Deep Learning for robot control, this paper considers two learning algorithms in terms of how they acquire demonstrations from fallible human supervisors. Human-Centric (HC) sampling is a standard supervised learning algorithm, where a human supervisor demonstrates the task by teleoperating the robot to provide trajectories consisting of state-control pairs. Robot-Centric (RC) sampling is an increasingly popular alternative used in algorithms such as DAgger, where a human supervisor observes the robot execute a learned policy and provides corrective control labels for each state visited. We suggest RC sampling can be challenging for human supervisors and prone to mislabeling. RC sampling can also induce error in policy performance because it repeatedly visits areas of the state space that are harder to learn. Although policies learned with RC sampling can be superior to HC sampling for standard learning models such as linear SVMs, policies learned with HC sampling may be comparable to RC when applied to expressive learning models such as deep learning and hyper-parametric decision trees, which can achieve very low training error provided there is enough data. We compare HC and RC using a grid world environment and a physical robot singulation task. In the latter the input is a binary image of objects on a planar worksurface and the policy generates a motion in the gripper to separate one object from the rest. We observe in simulation that for linear SVMs, policies learned with RC outperformed those learned with HC but that using deep models this advantage disappears. We also find that with RC, the corrective control labels provided by humans can be highly inconsistent. We prove there exists a class of examples in which at the limit, HC is guaranteed to converge to an optimal policy while RC may fail to converge. These results suggest a form of HC sampling may be preferable for highly-expressive learning models and human supervisors.