Chain-of-thought reasoning, where language models expend additional computation by producing thinking tokens prior to final responses, has driven significant advances in model capabilities. However, training these reasoning models is extremely costly in terms of both data and compute, as it involves collecting long traces of reasoning behavior from humans or synthetic generators and further post-training the model via reinforcement learning. Are these costs fundamental, or can they be reduced through better algorithmic design? We show that \textit{autocurriculum}—where the model uses its own performance to decide which problems to focus training on—provably improves upon standard training recipes for both supervised fine-tuning (SFT) and reinforcement learning (RL). For SFT, we show that autocurriculum requires \textit{exponentially} fewer reasoning demonstrations than non-adaptive fine-tuning (Joshi et al., 2025), by focusing teacher supervision on prompts where the current model struggles. For RL fine-tuning, autocurriculum \textit{decouples} the computational cost from the quality of the reference model, reducing the latter to a burn-in cost that is nearly independent of the target accuracy. These improvements arise purely from adaptive data selection, drawing on classical techniques from boosting (Freund and Schapire, 1997) and learning from counterexamples (Angluin, 1987), and requiring no assumption on the distribution or difficulty of prompts.

Next-token prediction with the logarithmic loss is a cornerstone of autoregressive sequence modeling, but, in practice, suffers from \emph{error amplification}, where errors in the model compound and generation quality degrades as sequence length $H$ increases. From a theoretical perspective, this phenomenon should not appear in \emph{well-specified} settings, and, indeed, a growing body of empirical work hypothesizes that \emph{misspecification}, where the learner is not sufficiently expressive to represent the target distribution, may be the root cause. Under misspecification—where the goal is to learn as well as the best-in-class model up to a multiplicative approximation factor $C\geq{}1$—we confirm that $C$ indeed grows with $H$ for next-token prediction, lending theoretical support to this empirical hypothesis. We then ask whether this mode of error amplification is avoidable algorithmically, computationally, or information-theoretically, and uncover inherent computational-statistical tradeoffs. We show: \textbf{(1)} Information-theoretically, one can avoid error amplification and achieve $C=O(1)$. \textbf{(2)} Next-token prediction can be made robust to achieve $C=\tilde{O}(H)$, representing moderate error amplification, but this is an inherent barrier: \emph{any} next-token prediction-style objective must suffer $C=\Omega(H)$. \textbf{(3)} For the natural testbed of autoregressive \emph{linear} models, \emph{no computationally efficient algorithm} can achieve sub-polynomial approximation factor $C=e^{(\log H)^{1-\Omega(1)}}$; however, at least for binary token spaces, one can smoothly trade compute for statistical power and improve on $C=\Omega(H)$ in sub-exponential time. Our results have consequences in the more general setting of imitation learning, where the widely-used behavior cloning generalizes next-token prediction.

Recent work in language modeling has raised the possibility of “self-improvement,” where an LLM evaluates and refines its own generations to achieve higher performance without external feedback. It is impossible for this self-improvement to create information that is not already in the model, so why should we expect that this will lead to improved capabilities? We offer a new theoretical perspective on the capabilities of self-improvement through a lens we refer to as “sharpening.” Motivated by the observation that language models are often better at verifying response quality than they are at generating correct responses, we formalize self-improvement as using the model itself as a verifier during post-training in order to ‘sharpen’ the model to one placing large mass on high-quality sequences, thereby amortizing the expensive inference-time computation of generating good sequences. We begin by introducing a new statistical framework for sharpening in which the learner has sample access to a pre-trained base policy. Then, we analyze two natural families of self improvement algorithms based on SFT and RLHF. We find that (i) the SFT-based approach is minimax optimal whenever the initial model has sufficient coverage, but (ii) the RLHF-based approach can improve over SFT-based self- improvement by leveraging online exploration, bypassing the need for coverage. We view these findings as a starting point toward a foundational understanding that can guide the design and evaluation of self-improvement algorithms.

Language model alignment methods such as reinforcement learning from human feedback (RLHF) have led to impressive advances in language model capabilities, but are limited by a widely observed phenomenon known as *overoptimization*, where the quality of the language model degrades over the course of the alignment process. As the model optimizes performance on an offline reward model, it overfits to inaccuracies and drifts away from preferred responses covered by the data. To discourage such distribution shift, KL-regularization is widely employed in existing offline alignment methods, but overoptimization continues to harm performance. Lending theoretical insight into the source of these empirical observations, we first show that the KL-regularization is too weak to prevent overfitting, then ask: is it possible to design an efficient algorithm that is provably robust to overoptimization? In this paper, we advance theoretical understanding of sample-efficient offline alignment and introduce a new algorithm called $\chi^2$-Preference Optimization ($\chi$PO). $\chi$PO is a one-line change to Direct Preference Optimization (DPO; Rafailov et al. 2023), that modifies only the logarithmic link function in the DPO objective. Despite this minimal change, $\chi$PO implicitly implements the principle of *pessimism in the face of uncertainty* via regularization with the $\chi^2$-divergence---which quantifies uncertainty more effectively than KL-regularization---and provably alleviates overoptimization, achieving sample-complexity guarantees based on *single-policy concentrability*, the gold standard in offline reinforcement learning. This guarantee makes $\chi$PO the first simple, yet general-purpose offline alignment algorithm that is provably robust to overoptimization.

Recent work on inference-time alignment has established benefits of increasing inference-time computation in language models, but naively scaling compute through techniques like Best-of-N sampling can cause performance to degrade due to reward hacking. Toward a theoretical understanding of how to best leverage additional computation, we formalize inference-time alignment as improving a pre-trained policy’s responses for a prompt of interest, given access to an imperfect reward model. We analyze the performance of inference-time alignment algorithms in terms of (i) response quality, and (ii) compute, and provide new results that highlight the importance of the pre-trained policy’s coverage over high-quality responses for performance and compute scaling: (1) We show that Best-of-N alignment with an ideal N can achieve optimal performance under stringent notions of coverage, but provably suffers from reward hacking when N is large, and fails to achieve tight guarantees under more realistic coverage conditions; (2) We introduce InferenceTimePessimism, a new algorithm which mitigates reward hacking through deliberate use of inference-time compute, implementing pessimism in the face of uncertainty; we prove that its performance is optimal and scaling-monotonic, i.e., does ot degrade as N increases. We complement our theoretical results with experiments that demonstrate the practicality of our algorithm across a variety of tasks and models.

Holdout validation and hyperparameter tuning from data is a long-standing problem in offline reinforcement learning (RL). A standard framework is to use off-policy evaluation (OPE) methods to evaluate and select the policies, but OPE either incurs exponential variance (e.g., importance sampling) or has hyperparameters on their own (e.g., FQE and model-based). We focus on hyperparameter tuning for OPE itself, which is even more under-investigated. Concretely, we select among candidate value functions ("model-free") or dynamics models ("model-based") to best assess the performance of a target policy. We develop: (1) new model-free and model-based selectors with theoretical guarantees, and (2) a new experimental protocol for empirically evaluating them. Compared to the model-free protocol in prior works, our new protocol allows for more stable generation and better control of candidate value functions in an optimization-free manner, and evaluation of model-free and model-based methods alike. We exemplify the protocol on Gym-Hopper, and find that our new model-free selector, LSTD-Tournament, demonstrates promising empirical performance.

In standard reinforcement learning setups, the agent receives observations and performs actions at evenly spaced intervals. However, in many real-world settings, observations are expensive, forcing agents to commit to courses of action for designated periods of time. Consider that doctors, after each visit, typically set not only a treatment plan but also a follow-up date at which that plan might be revised. In this work, we formalize the setup of timing-as-an-action. Through theoretical analysis in the tabular setting, we show that while the choice of delay intervals could be naively folded in as part of a composite action, these actions have a special structure and handling them intelligently yields statistical advantages. Taking a model-based perspective, these gains owe to the fact that delay actions do not add any parameters to the underlying model. For model estimation, we provide provable sample-efficiency improvements, and our experiments demonstrate empirical improvements in both healthcare simulators and classical reinforcement learning environments.

Occupancy functions play an instrumental role in reinforcement learning (RL) for guiding exploration, handling distribution shift, and optimizing general objectives beyond the expected return. Yet, computationally efficient policy optimization methods that use (only) occupancy functions are virtually non-existent. In this paper, we establish the theoretical foundations of model-free policy gradient (PG) methods that compute the gradient through the occupancy for both online and offline RL, without modeling value functions. Our algorithms reduce gradient estimation to squared-loss regression and are computationally oracle-efficient. We characterize the sample complexities of both local and global convergence, accounting for both finite-sample estimation error and the roles of exploration (online) and data coverage (offline). Occupancy-based PG naturally handles arbitrary offline data distributions, and, with one-line algorithmic changes, can be adapted to optimize any differentiable objective functional.

MDPs with low-rank transitions---that is, the transition matrix can be factored into the product of two matrices, left and right---is a highly representative structure that enables tractable learning. The left matrix enables expressive function approximation for value-based learning and has been studied extensively. In this work, we instead investigate sample-efficient learning with density features, i.e., the right matrix, which induce powerful models for state-occupancy distributions. This setting not only sheds light on leveraging unsupervised learning in RL, but also enables plug-in solutions for settings like convex RL. In the offline setting, we propose an algorithm for off-policy estimation of occupancies that can handle non-exploratory data. Using this as a subroutine, we further devise an online algorithm that constructs exploratory data distributions in a level-by-level manner. As a central technical challenge, the additive error of occupancy estimation is incompatible with the multiplicative definition of data coverage. In the absence of strong assumptions like reachability, this incompatibility easily leads to exponential error blow-up, which we overcome via novel technical tools. Our results also readily extend to the representation learning setting, when the density features are unknown and must be learned from an exponentially large candidate set.

Addressing such diverse ends as mitigating safety risks, aligning agent behavior with human preferences, and improving the efficiency of learning, an emerging line of reinforcement learning research addresses the entire distribution of returns and various risk functionals that depend upon it. In the contextual bandit setting, recently work on off-policy risk assessment estimates the target policy’s CDF of returns, providing finite sample guarantees that extend to (and hold simultaneously over) plugin estimates of an arbitrarily large set of risk functionals. In this paper, we lift OPRA to Markov decision processes (MDPs), where importance sampling (IS) CDF estimators suffer high variance on longer trajectories due to vanishing (and exploding) importance weights. To mitigate these problems, we incorporate model-based estimation to develop the first doubly robust (DR) estimator for the CDF of returns in MDPs. The DR estimator enjoys significantly less variance and, when the model is well specified, achieves the Cramer-Rao variance lower bound. Moreover, for many risk functionals, the downstream estimates enjoy both lower bias and lower variance. Additionally, we derive the first minimax lower bounds for off-policy CDF and risk estimation, which match our error bounds up to a constant. Finally, we demonstrate the efficacy of our DR CDF estimates experimentally on several different environments.

Sample-efficiency guarantees for offline reinforcement learning (RL) often rely on strong assumptions on both the function classes (e.g., Bellman-completeness) and the data coverage (e.g., all-policy concentrability). Despite the recent efforts on relaxing these assumptions, existing works are only able to relax one of the two factors, leaving the strong assumption on the other factor intact. As an important open problem, can we achieve sample-efficient offline RL with weak assumptions on both factors? In this paper we answer the question in the positive. We analyze a simple algorithm based on the primal-dual formulation of MDPs, where the dual variables (discounted occupancy) are modeled using a density-ratio function against offline data. With proper regularization, the algorithm enjoys polynomial sample complexity, under only realizability and single-policy concentrability. We also provide alternative analyses based on different assumptions to shed light on the nature of primal-dual algorithms for offline RL.

Standard uniform convergence results bound the generalization gap of the expected loss over a hypothesis class. The emergence of risk-sensitive learning requires generalization guarantees for functionals of the loss distribution beyond the expectation. While prior works specialize in uniform convergence of particular functionals, our work provides uniform convergence for a general class of Hölder risk functionals for which the closeness in the Cumulative Distribution Function (CDF) entails closeness in risk. We establish the first uniform convergence results for estimating the CDF of the loss distribution, which yield uniform convergence guarantees that hold simultaneously both over a class of Hölder risk functionals and over a hypothesis class. Thus licensed to perform empirical risk minimization, we develop practical gradient-based methods for minimizing distortion risks (widely studied subset of Hölder risks that subsumes the spectral risks, including the mean, conditional value at risk, cumulative prospect theory risks, and others) and provide convergence guarantees. In experiments, we demonstrate the efficacy of our learning procedure, both in settings where uniform convergence results hold and in high-dimensional settings with deep networks.

Off-policy evaluation often refers to two related tasks: estimating the expected return of a policy and estimating its value function (or other functions of interest, such as density ratios). While recent works on marginalized importance sampling (MIS) show that the former can enjoy provable guarantees under realizable function approximation, the latter is only known to be feasible under much stronger assumptions such as prohibitively expressive discriminators. In this work, we provide guarantees for off-policy function estimation under only realizability, by imposing proper regularization on the MIS objectives. Compared to commonly used regularization in MIS, our regularizer is much more flexible and can account for an arbitrary user-specified distribution, under which the learned function will be close to the groundtruth. We provide exact characterization of the optimal dual solution that needs to be realized by the discriminator class, which determines the data-coverage assumption in the case of value-function learning. As another surprising observation, the regularizer can be altered to relax the data-coverage requirement, and completely eliminate it in the ideal case with strong side information.

Even when unable to run experiments, practitioners can evaluate prospective policies, using previously logged data. However, while the bandits literature has adopted a diverse set of objectives, most research on off-policy evaluation to date focuses on the expected reward. In this paper, we introduce Lipschitz risk functionals, a broad class of objectives that subsumes conditional value-at-risk (CVaR), variance, mean-variance, many distorted risks, and CPT risks, among others. We propose Off-Policy Risk Assessment (OPRA), a framework that first estimates a target policy's CDF and then generates plugin estimates for any collection of Lipschitz risks, providing finite sample guarantees that hold simultaneously over the entire class. We instantiate OPRA with both importance sampling and doubly robust estimators. Our primary theoretical contributions are (i) the first uniform concentration inequalities for both CDF estimators in contextual bandits and (ii) error bounds on our Lipschitz risk estimates, which all converge at a rate of $O(1/\sqrt{n})$.