When developing new large language models (LLMs), a key step is evaluating their final performance, often by computing the win-rate against a reference model based on external feedback. Human feedback is the gold standard, particularly for capturing nuanced qualities like coherence, readability, and alignment with human expectations. However, human evaluations are costly—even for large tech companies—and when conducted with active users, they may negatively impact user experience. A promising alternative is synthetic feedback, where evaluations are conducted by other large language models, including reward models. While this eliminates the need for costly human annotations, it introduces biases that may distort the evaluation process. In this work, we propose a statistically principled framework that integrates human and synthetic feedback to reduce reliance on human annotations while maintaining unbiased win-rate calculations. Our experiments demonstrate a reduction in human annotations by up to 12.2\% with an off-the-shelf synthetic evaluator and up to 24.8\% with a finetuned variant. Apart from being generalizable, scalable, and free of hyper-parameter tuning, our method offers predictable annotation savings, which can be estimated based on data-dependent characteristics.

Scaling model size and training data has led to great advances in the performance of Large Language Models (LLMs). However, the diminishing returns of this approach necessitate alternative methods to improve model capabilities, particularly in tasks requiring advanced reasoning. Large reasoning models, which leverage long chain-of-thoughts, bring unprecedented breakthroughs in problem-solving capabilities but at a substantial deployment cost associated to longer generations. Reducing inference costs is crucial for the economic feasibility, user experience, and environmental sustainability of these models. In this work, we propose to train large reasoning models to reason efficiently. Our method incentivizes models to minimize unnecessary computational overhead while largely maintaining accuracy, thereby achieving substantial deployment efficiency gains. It enables the derivation of a family of reasoning models with varying efficiency levels, controlled via a single hyperparameter. Experiments on two open-weight large reasoning models demonstrate significant reductions in inference cost while preserving most of the accuracy.

Large language models (LLMs) have the potential to tackle sequential decision-making problems due to their generalist capabilities. Instead of optimizing ``myopic'' surrogate objectives such as human preferences within a single turn, in such problems, we wish to directly optimize long-term objectives, such as user satisfaction over an entire dialogue with an LLM or delayed success metrics in web navigation. Multi-turn reinforcement learning (RL) provides an appealing approach to directly optimize long-term objectives, but how can we design effective and efficient multi-turn RL algorithms for LLMs? In this work, we propose an algorithmic framework to multi-turn RL for LLMs that preserves the flexibility of token-by-token RL used in single-turn RL problems, while still accommodating long horizons and delayed rewards more effectively. Our framework, the **A**cto**r**-**C**ritic Framework with a **H**i**e**rarchical Structu**r**e (**ArCHer**), combines a high-level off-policy RL algorithm that trains a value function with a low-level RL algorithm that trains a token-by-token policy. While ArCHer can be instantiated with multiple RL algorithms, a particularly convenient instantiation is to use temporal difference (TD) learning at the high level and on-policy token-level policy gradient at the low level. Empirically, we show that ArCHer significantly improves efficiency and performance of multi-turn LLM tasks, attaining sample efficiency boosts of about **100x** over prior on-policy methods and converging to a much better performance than other off-policy methods.

The safe and effective deployment of Large Language Models (LLMs) involves a critical step called alignment, which ensures that the model's responses are in accordance with human preferences. Prevalent alignment techniques, such as DPO, PPO and their variants, align LLMs by changing the pre-trained model weights during a phase called post-training. While predominant, these post-training methods add substantial complexity before LLMs can be deployed. Inference-time alignment methods avoid the complex post-training step and instead bias the generation towards responses that are aligned with human preferences. The best-known inference-time alignment method, called Best-of-N, is as effective as the state-of-the-art post-training procedures. Unfortunately, Best-of-N requires vastly more resources at inference time than standard decoding strategies, which makes it computationally not viable. In this work, we introduce Speculative Rejection, a computationally-viable inference-time alignment algorithm. It generates high-scoring responses according to a given reward model, like Best-of-N does, while being between 16 to 32 times more computationally efficient.

Understanding when reinforcement learning algorithms can make successful off-policy predictions—and when the may fail to do so–remains an open problem. Typically, model-free algorithms for reinforcement learning are analyzed under a condition called Bellman completeness when they operate off-policy with function approximation, unless additional conditions are met. However, Bellman completeness is a requirement that is much stronger than realizability and that is deemed to be too strong to hold in practice. In this work, we relax this structural assumption and analyze the statistical complexity of off-policy reinforcement learning when only realizability holds for the prescribed function class. We establish finite-sample guarantees for off-policy reinforcement learning that are free of the approximation error term known as inherent Bellman error, and that depend on the interplay of three factors. The first two are well known: they are the metric entropy of the function class and the concentrability coefficient that represents the cost of learning off-policy. The third factor is new, and it measures the violation of Bellman completeness, namely the mis-alignment between the chosen function class and its image through the Bellman operator. Our analysis directly applies to the solution found by temporal difference algorithms when they converge.

In some applications of reinforcement learning, a dataset of pre-collected experience is already available but it is also possible to acquire some additional online data to help improve the quality of the policy. However, it may be preferable to gather additional data with a single, non-reactive exploration policy and avoid the engineering costs associated with switching policies. In this paper we propose an algorithm with provable guarantees that can leverage an offline dataset to design a single non-reactive policy for exploration. We theoretically analyze the algorithm and measure the quality of the final policy as a function of the local coverage of the original dataset and the amount of additional data collected.

The Q-learning algorithm is a simple, fundamental and practically very effective reinforcement learning algorithm. However, the basic protocol can exhibit an unstable behavior when implemented even with simple linear function approximation. While tools like target networks and experience replay are often implemented to stabilize the learning process, the individual contribution of each of these mechanisms is not well understood theoretically. This work proposes an exploration variant of the basic Q-learning protocol with linear function approximation. Our modular analysis illustrates the role played by each algorithmic tool that we adopt: a second order update rule, a set of target networks, and a mechanism akin to experience replay. Together, they enable state of the art regret bounds on linear MDPs while preserving the most prominent feature of the algorithm, namely a space complexity independent of the number of steps elapsed. Furthermore, we show that the performance of the algorithm degrades very gracefully under a new, more permissive notion of approximation error. Finally, the algorithm partially inherits problem dependent regret bounds, function of the number of ‘effective’ feature dimension.

We propose and analyze a reinforcement learning principle thatapproximates the Bellman equations by enforcing their validity onlyalong a user-defined space of test functions. Focusing onapplications to model-free offline RL with function approximation, weexploit this principle to derive confidence intervals for off-policyevaluation, as well as to optimize over policies within a prescribedpolicy class. We prove an oracle inequality on our policyoptimization procedure in terms of a trade-off between the value anduncertainty of an arbitrary comparator policy. Different choices oftest function spaces allow us to tackle different problems within acommon framework. We characterize the loss of efficiency in movingfrom on-policy to off-policy data using our procedures, and establishconnections to concentrability coefficients studied in past work. Weexamine in depth the implementation of our methods with linearfunction approximation, and provide theoretical guarantees withpolynomial-time implementations even when Bellman closure does nothold.

Policy optimization methods are popular reinforcement learning (RL) algorithms, because their incremental and on-policy nature makes them more stable than the value-based counterparts. However, the same properties also make them slow to converge and sample inefficient, as the on-policy nature is not amenable to data reuse and the incremental updates result in a large iteration complexity before a near optimal policy is discovered. These empirical findings are mirrored in theory in the recent work of \citet{agarwal2020pc}, which provides a policy optimization method PC-PG that provably finds near optimal policies in the linear MDP model and is robust to model misspecification, but suffers from an extremely poor sample complexity compared with value-based techniques. In this paper, we propose a new algorithm, COPOE, that overcomes the poor sample complexity of PC-PG while retaining the robustness to model misspecification. COPOE makes several important algorithmic enhancements of PC-PG, such as enabling data reuse, and uses more refined analysis techniques, which we expect to be more broadly applicable to designing new RL algorithms. The result is an improvement in sample complexity from $\widetilde{O}(1/\epsilon^{11})$ for PC-PG to $\widetilde{O}(1/\epsilon^3)$ for COPOE, nearly bridging the gap with value-based techniques.

Several practical applications of reinforcement learning involve an agent learning from past data without the possibility of further exploration. Often these applications require us to 1) identify a near optimal policy or to 2) estimate the value of a target policy. For both tasks we derive exponential information-theoretic lower bounds in discounted infinite horizon MDPs with a linear function representation for the action value function even if 1) realizability holds, 2) the batch algorithm observes the exact reward and transition functions, and 3) the batch algorithm is given the best a priori data distribution for the problem class. Our work introduces a new ‘oracle + batch algorithm’ framework to prove lower bounds that hold for every distribution. The work shows an exponential separation between batch and online reinforcement learning.

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.

Actor-critic methods are widely used in offline reinforcement learningpractice, but are not so well-understood theoretically. We propose a newoffline actor-critic algorithm that naturally incorporates the pessimism principle, leading to several key advantages compared to the state of the art. The algorithm can operate when the Bellman evaluation operator is closed with respect to the action value function of the actor's policies; this is a more general setting than the low-rank MDP model. Despite the added generality, the procedure is computationally tractable as it involves the solution of a sequence of second-order programs.We prove an upper bound on the suboptimality gap of the policy returned by the procedure that depends on the data coverage of any arbitrary, possibly data dependent comparator policy.The achievable guarantee is complemented with a minimax lower bound that is matching up to logarithmic factors.

We consider the exploration-exploitation dilemma in finite-horizon reinforcement learning (RL). When the state space is large or continuous, traditional tabular approaches are unfeasible and some form of function approximation is mandatory. In this paper, we introduce an optimistically-initialized variant of the popular randomized least-squares value iteration (RLSVI), a model-free algorithm where exploration is induced by perturbing the least-squares approximation of the action-value function. Under the assumption that the Markov decision process has low-rank transition dynamics, we prove that the frequentist regret of RLSVI is upper-bounded by $\widetilde O(d^2 H^2 \sqrt{T})$ where $ d $ are the feature dimension, $ H $ is the horizon, and $ T $ is the total number of steps. To the best of our knowledge, this is the first frequentist regret analysis for randomized exploration with function approximation.

We study the exploration problem with approximate linear action-value functions in episodic reinforcement learning under the notion of low inherent Bellman error, a condition normally employed to show convergence of approximate value iteration. First we relate this condition to other common frameworks and show that it is strictly more general than the low rank (or linear) MDP assumption of prior work. Second we provide an algorithm with a high probability regret bound $\widetilde O(\sum_{t=1}^H d_t \sqrt{K} + \sum_{t=1}^H \sqrt{d_t} \IBE K)$ where $H$ is the horizon, $K$ is the number of episodes, $\IBE$ is the value if the inherent Bellman error and $d_t$ is the feature dimension at timestep $t$. In addition, we show that the result is unimprovable beyond constants and logs by showing a matching lower bound. This has two important consequences: 1) it shows that exploration is possible using only \emph{batch assumptions} with an algorithm that achieves the optimal statistical rate for the setting we consider, which is more general than prior work on low-rank MDPs 2) the lack of closedness (measured by the inherent Bellman error) is only amplified by $\sqrt{d_t}$ despite working in the online setting. Finally, the algorithm reduces to the celebrated \textsc{LinUCB} when $H=1$ but with a different choice of the exploration parameter that allows handling misspecified contextual linear bandits. While computational tractability questions remain open for the MDP setting, this enriches the class of MDPs with a linear representation for the action-value function where statistically efficient reinforcement learning is possible.

There has been growing progress on theoretical analyses for provably efficient learning in MDPs with linear function approximation, but much of the existing work has made strong assumptions to enable exploration by conventional exploration frameworks. Typically these assumptions are stronger than what is needed to find good solutions in the batch setting. In this work, we show how under a more standard notion of low inherent Bellman error, typically employed in least-square value iteration-style algorithms, we can provide strong PAC guarantees on learning a near optimal value function provided that the linear space is sufficiently ``explorable''. We present a computationally tractable algorithm for the reward-free setting and show how it can be used to learn a near optimal policy for any (linear) reward function, which is revealed only once learning has completed. If this reward function is also estimated from the samples gathered during pure exploration, our results also provide same-order PAC guarantees on the performance of the resulting policy for this setting.

Strong worst-case performance bounds for episodic reinforcement learning exist but fortunately in practice RL algorithms perform much better than such bounds would predict. Algorithms and theory that provide strong problem-dependent bounds could help illuminate the key features of what makes a RL problem hard and reduce the barrier to using RL algorithms in practice. As a step towards this we derive an algorithm and analysis for finite horizon discrete MDPs with state-of-the-art worst-case regret bounds and substantially tighter bounds if the RL environment has special features but without apriori knowledge of the environment from the algorithm. As a result of our analysis, we also help address an open learning theory question \cite{jiang2018open} about episodic MDPs with a constant upper-bound on the sum of rewards, providing a regret bound function of the number of episodes with no dependence on the horizon.

This paper focuses on the problem of computing an $\epsilon$-optimal policy in a discounted Markov Decision Process (MDP) provided that we can access the reward and transition function through a generative model. We propose an algorithm that is initially agnostic to the MDP but that can leverage the specific MDP structure, expressed in terms of variances of the rewards and next-state value function, and gaps in the optimal action-value function to reduce the sample complexity needed to find a good policy, precisely highlighting the contribution of each state-action pair to the final sample complexity. A key feature of our analysis is that it removes all horizon dependencies in the sample complexity of suboptimal actions except for the intrinsic scaling of the value function and a constant additive term.

We study linear approximate value iteration (LAVI) with a generative model. While linear models may accurately represent the optimal value function using a few parameters, several empirical and theoretical studies show the combination of least-squares projection with the Bellman operator may be expansive, thus leading LAVI to amplify errors over iterations and eventually diverge. We introduce an algorithm that approximates value functions by combining Q-values estimated at a set of \textit{anchor} states. Our algorithm tries to balance the generalization and compactness of linear methods with the small amplification of errors typical of interpolation methods. We prove that if the features at any state can be represented as a convex combination of features at the anchor points, then errors are propagated linearly over iterations (instead of exponentially) and our method achieves a polynomial sample complexity bound in the horizon and the number of anchor points. These findings are confirmed in preliminary simulations in a number of simple problems where a traditional least-square LAVI method diverges.

In order to make good decision under uncertainty an agent must learn from observations. To do so, two of the most common frameworks are Contextual Bandits and Markov Decision Processes (MDPs). In this paper, we study whether there exist algorithms for the more general framework (MDP) which automatically provide the best performance bounds for the specific problem at hand without user intervention and without modifying the algorithm. In particular, it is found that a very minor variant of a recently proposed reinforcement learning algorithm for MDPs already matches the best possible regret bound $\tilde O (\sqrt{SAT})$ in the dominant term if deployed on a tabular Contextual Bandit problem despite the agent being agnostic to such setting.