Offline-to-Online Contextual Bandits

Long-term memory enables language-model agents to personalize future interactions, but a finite memory bank cannot preserve every extracted fact. Existing memory gating policies are typically shared across users and therefore fail to account for user-specific patterns of future reuse. In this paper, we study memory gating at the fact level. We first propose a retrieval-based labeling strategy that labels each fact by whether it is later retrieved and incorporated into responses. We then formulate online memory gating as contextual apple tasting with delayed one-sided feedback, and propose Warm-Start UCB, which combines offline learning from existing data with online adaptation to the current user. Our theoretical analysis establishes a high-probability regret upper bound of $\tilde{O}(Tε\wedge d\sqrt{T})$ when the delay window is at most $O(\sqrt{T})$, where $T$ is the total number of rounds, $ε$ is the initialization error, $d$ is the dimension of the feature map, and $\tilde{O}(\cdot)$ hides constant and poly-logarithmic factors. The results of our experiment demonstrate that our method outperforms all baselines under different memory budgets and history lengths, establishing it as a promising personalized memory gating policy.

Pairwise losses are increasingly used for reward learning even when pointwise rewards are observed, with mixed empirical results. When and why do pairwise losses outperform pointwise losses? We study this question in a grouped offline contextual-bandit setting allowing multiple actions per context, capturing many reward learning scenarios. We compare Value Regression (VR), which regresses observed rewards pointwise, with Value Difference Regression (VDR), which regresses reward differences between a pair of actions sampled under the same context. We consider a semiparametric model where the mean reward is the sum of a learnable action-dependent component and an arbitrary context-dependent yet action-independent nuisance, capturing context-specific disturbances. Using a unified localized analysis, we prove finite-sample regression guarantees for finite and linear function classes and translate them into offline-regret bounds. For finite classes, VDR eliminates the misspecification term in the VR bound and improves a reward-scale-dependent error term by averaging over actions within each context, a benefit absent from the corresponding VR term. For linear classes, neither method uniformly dominates: within-context differencing removes nuisance-induced bias but may increase estimation variance relative to using absolute rewards when the misspecification is sufficiently low. This yields a feature geometry-dependent bias-variance tradeoff, which we corroborate with numerical experiments.

Personalizing marketing messages with contextual multi-armed bandits (CMABs) drives real business value, yet the objective that ultimately matters - a downstream conversion - is observed only weeks later, too late to drive online learning. Teams therefore train the bandit on a fast proxy reward, and separately must judge whether a contextual bandit is worth its complexity over sending one best message. Settling both decisions with the usual offline checks - a batch off-policy estimate, a marginal arm-discrimination test, a confidence interval - can mislead systematically under delayed feedback. We give an ordered diagnostic protocol that screens a reward-and-policy candidate on two axes, alignment (does optimizing the reward move the north-star?) and learnability (can the bandit identify the reward-optimal policy?), before trusting any reported lift. We validate it where the truth is known - a public off-policy-evaluation benchmark and a controllable synthetic generator - and illustrate it on a deployed large-marketplace push system (where, with five arms and one split, the evidence is directional rather than powered). Two lessons recur. (N1) A single offline number can mis-rank rewards: a denser reward signal gives the bandit more to learn from, so rewards that look tied in a static estimate pull apart once learning happens online. (N2) If you cannot tell in advance which single message is best, a per-user policy partly just avoids betting on the wrong one - that looks like personalization but is really robustness, so a "personalization premium" is easily overstated. Our contribution is methodological rather than algorithmic: the ordered protocol, the two lessons it surfaces, and the end-to-end experience of applying it to a delayed-feedback CMAB.

Optimizing instructional policies in Intelligent Tutoring Systems (ITS) typically requires costly online experimentation or student simulators that may fail to capture real-world dynamics. This paper introduces an offline contextual bandit framework that learns new adaptive policies directly from logged interaction data. By mapping student-item interactions onto a continuous latent proficiency-difficulty scale using a Rasch model, we cast the tutoring process as a continuous stochastic bandit problem. We propose a novel reward function designed to optimize ''flow'' by balancing task challenge with student success. Our approach includes a round-specific behavior policy estimation that serves as both a propensity model for off-policy evaluation and a diagnostic tool for ITS adaptivity. We demonstrate the efficacy of this framework across four large-scale real-world datasets, achieving consistent policy improvements over the logged behavior policy. The results show that effective instructional policies can be learned and visualized within seconds of computation, providing a scalable path for improving adaptive learning systems without further data collection.

The rapid development of large language models, each with distinct capabilities and inference costs, raises a practical deployment question: given an incoming request, which model should handle it? We present OrcaRouter, a production-oriented LLM router that combines a LinUCB-based contextual bandit over lexical and sentence-embedding features with a hybrid offline-online learning protocol. Offline, OrcaRouter obtains full-information feedback by evaluating each candidate model on a curated set of routing prompts, yielding a reward matrix used to fit one ridge regressor per arm. At deployment time, it initializes from these parameters and can optionally continue learning from bandit feedback, updating only the selected model's arm after observing its reward. At the time of our RouterArena submission (May 20, 2026), OrcaRouter-Adaptive ranked second on the public RouterArena leaderboard with an arena score of 72.08, achieving 75.54% accuracy at a cost of USD 1.00 per 1,000 queries.

Automated decision-making algorithms drive applications such as recommendation systems and search engines. These algorithms often rely on off-policy contextual bandits or off-policy learning (OPL). Conventionally, OPL selects actions that maximize the expected reward from an existing action set. However, in many real-world scenarios, actions, such as news articles or video content, change continuously, and the action space evolves over time after data collection. We define actions introduced after deploying the logging policy as new actions and focus on OPL with new actions. Existing OPL methods identify optimal actions from the existing set effectively but cannot learn and select new actions because no relevant data are logged. To address this limitation, we propose a new OPL method that leverages action features. We first introduce the Local Combination PseudoInverse (LCPI) estimator for the policy gradient, generalizing the PseudoInverse estimator initially proposed for off-policy evaluation of slate bandits. LCPI controls the trade-off between reward-modeling condition and the condition for data collection regarding the action features, capturing the interaction effects among different dimensions of action features. Furthermore, we propose a generalized algorithm called Policy Optimization for Effective New Actions (PONA), which integrates LCPI, a component specialized for new action selection, with Doubly Robust (DR), which excels at learning within existing actions. We define PONA as a weighted sum of the LCPI and DR estimators, optimizing both the selection of existing and new actions, and allowing the proportion of new action selections to be adjusted by the weight parameter. Through extensive experiments, we demonstrate that PONA efficiently selects new actions while maintaining the overall policy performance as opposed to most existing methods that cannot select new actions.

We study risk-aware offline policy learning, aiming to learn a decision rule from logged data that is optimal under general risk criteria. This problem is crucial in high-stakes domains where online interaction is infeasible and adverse outcomes must be carefully controlled. However, existing literature on offline contextual bandits either centers on expected-reward criteria or restricts risk considerations to policy evaluation instead of optimization. In this work, we propose a unified distributional framework for optimizing Lipschitz-continuous risk functionals, a broad class of risk measures encompassing mean-variance, entropic risk, and conditional value-at-risk, among others. By developing novel empirical concentration inequalities for importance sampling-based distributional estimators, our analysis derives data-dependent suboptimality bounds with an $\tilde{\mathcal{O}}(1/\sqrt{n})$ rate, without relying on restrictive uniform overlap assumptions. This rate is minimax optimal and matches that of risk-neutral offline policy optimization, indicating that optimizing general Lipschitz risk criteria incurs no additional statistical cost relative to the expected-reward.

Offline-to-online learning aims to improve online decision-making by leveraging offline logged data. A central challenge in this setting is the distribution shift between offline and online environments. While some existing works attempt to leverage shifted offline data, they largely rely on UCB-type algorithms. Thompson sampling (TS) represents another canonical class of bandit algorithms, well known for its strong empirical performance and naturally suited to offline-to-online learning through its Bayesian formulation. However, unlike UCB indices, posterior samples in TS are not guaranteed to be optimistic with respect to the true arm means. This makes indices constructed from purely online and hybrid data difficult to compare and complicates their use. To address this issue, we propose sample-mean anchored TS (Anchor-TS), which introduces a novel median-based anchoring rule that defines the arm index as the median of an online posterior sample, a hybrid posterior sample, and the online sample mean. The median anchoring systematically corrects bias induced by distribution shift by mitigating over-estimation for suboptimal arms and under-estimation for optimal arms, while exploiting offline information to obtain more accurate estimates when the shift is small. We establish theoretical guarantees showing that the proposed algorithm safely leverages offline data to accelerate online learning, and quantifying how the degree of distribution shift and the size of offline data affect the resulting regret reduction. Extensive experiments demonstrate consistent improvements of our algorithm over baselines.

Kullback-Leibler (KL) regularization is ubiquitous in reinforcement learning algorithms in the form of reverse or forward KL. Recent studies have demonstrated $ε^{-1}$-type fast rates for decision making under reverse KL regularization, in contrast to the standard $ε^{-2}$-type sample complexity. However, for forward-KL-regularized objectives, existing statistical analyses are either not applicable or result in $\tilde{O}(ε^{-2})$ slow rates. We take the first step towards addressing this problem via a streamlined analysis of forward-KL-regularized offline CBs. We give the first $\tilde{O}(ε^{-1})$ upper bounds in tabular and general function approximation settings, both under notions of single-policy concentrability. In particular, our convex-analytical pipeline unifies these settings by exploiting the pessimism principle in a novel way and completely bypasses the proof routines in previous works based on the mean value theorem, which might be of independent interest. Moreover, we provide rate-optimal lower bounds, manifesting the tightness of our upper bounds in terms of statistical rates. Our lower bounds also demonstrate that the forward-KL-regularized sample complexity recovers the unregularized slow rate in the low-regularization regime, similarly to the reverse-KL regularization.

Many bandit systems are deployed with offline historical data, such as past logs from earlier policies. Using these data can reduce early online exploration when they remain informative for the online problem. When the offline and online environments differ, such data can be biased for the online problem. For linear (contextual) bandits, this bias is directional: offline data may be informative in some feature directions and misleading in others. However, prior work typically controls this gap through a known Euclidean bound on the model parameters, which we prove is too coarse: even with the offline parameter known, bias in a single unknown direction can force dimension-dependent regret. To address this challenge, we introduce a directional bias certificate $(M_{\mathrm{bias}},ρ)$ that measures the offline-to-online gap through an $M_{\mathrm{bias}}$-induced norm and assigns different bias budgets to different directions. Building on this certificate, we propose Ellipsoidal-MINUCB, which augments the online learning with an offline-pooled branch that safely exploits historical data. When the certificate is known, we show that the algorithm matches the standard SupLinUCB rate in the worst case and improves when offline coverage aligns with low-bias directions. When the certificate is unknown, we estimate it adaptively from offline and accumulated online data and establish a corresponding regret guarantee. Numerical experiments support the theory and show gains in aligned regimes.

Active multi-target tracking requires a mobile robot to balance exploration for undetected targets with exploitation of uncertain tracked ones. Diffusion policies have emerged as a powerful approach for capturing diverse behavioral strategies by learning action sequences from expert demonstrations. However, existing methods implicitly select among strategies through the denoising process, without uncertainty quantification over which strategy to execute. We formulate expert selection for diffusion policies as an offline contextual bandit problem and propose a Bayesian framework for pessimistic, uncertainty-aware strategy selection. A multi-head Variational Bayesian Last Layer (VBLL) model predicts the expected tracking performance of each expert strategy given the current belief state, providing both a point estimate and predictive uncertainty. Following the pessimism principle for offline decision-making, a Lower Confidence Bound (LCB) criterion then selects the expert whose worst-case predicted performance is best, avoiding overcommitment to experts with unreliable predictions. The selected expert conditions a diffusion policy to generate corresponding action sequences. Experiments on simulated indoor tracking scenarios demonstrate that our approach outperforms both the base diffusion policy and standard gating methods, including Mixture-of-Experts selection and deterministic regression baselines.

We propose an algorithmic framework, Offline Estimation to Decisions (OE2D), that efficiently reduces contextual bandit learning with general reward function approximation to offline regression. The framework allows near-optimal regret for contextual bandits with large action spaces with $O(\log T)$ calls to an offline regression oracle over $T$ rounds, and makes $O(\log\log T)$ calls when $T$ is known. The design of OE2D algorithm generalizes Falcon~\citep{simchi2022bypassing} and its linear reward version~\citep[][Section 4]{xu2020upper} in that it finds an action distribution that we term ``exploitative F-design'' that simultaneously guarantees low regret and good coverage, striking a balance between exploration and exploitation. Central to our regret analysis is a new complexity measure, the Decision-Offline Estimation Coefficient (DOEC), which we show is small in many settings, including bounded Eluder dimension per-context and the smoothed regret setting. We also establish a relationship between DOEC and Decision Estimation Coefficient (DEC)~\citep{foster2021statistical}, bridging the design principles of offline- and online-oracle efficient contextual bandit algorithms for the first time.