The paper presents a framework for optimizing AI system configurations using Gaussian process bandits with selectively verified feedback. It introduces a fixed audit design that balances reliability coverage and reward separation, derives finite‑time regret bounds under a hard audit budget, and proposes a safe adaptive extension that mixes this design with Chernoff allocations. Experiments on synthetic data compare the certified designs to adaptive heuristics and standard GP controls, showing that safe adaptive allocation can reduce aggregate regret but may have loose finite‑budget error certificates.
arXiv:2601. 07094v2 Announce Type: replace-cross Abstract: Bayesian optimization (BO) iteratively fits a Gaussian process (GP) surrogate to accumulated evaluations and selects new queries via an acquisition function.
arXiv:2608. 08002v1 Announce Type: new Abstract: Language-model judges and reward models enable scalable supervision, but finite optimization can exploit evaluator errors rather than improve response quality.
The paper introduces a new approach to safety in contextual bandits with continuous actions by enforcing high‑probability constraints on the realized cost rather than on its expectation. It proposes the High‑Probability Constrained UCB algorithm, which balances optimistic reward exploration with pessimistic safety estimation. The authors provide theoretical regret guarantees for linear models and extend the analysis to general function classes, demonstrating experimentally that realized‑cost constraints significantly reduce safety violations compared to expected‑cost baselines.
The paper introduces a new approach to safety in contextual bandits with continuous actions, focusing on high‑probability constraints on the realized cost rather than expected cost. It presents the High‑Probability Constrained UCB algorithm, which balances reward exploration with conservative safety estimation, and provides theoretical regret guarantees for linear models and extensions to general function classes. Experiments demonstrate that this realized‑cost safety framework significantly reduces safety violations compared to expected‑cost constrained methods.
arXiv:2604. 26360v2 Announce Type: replace-cross Abstract: Reinforcement learning from human feedback (RLHF) systems face a compounding alignment challenge: not only are learned reward models uncertain about unseen state-action pairs, but the human preference annotations they are trained on are themselves inconsistent, context-dependent, and noisy.
arXiv:2608. 07725v1 Announce Type: new Abstract: Average-reward reinforcement-learning regret is known up to logarithmic factors, but the numerical content of published guarantees is difficult to compare because probability mode, structural parameter, logarithmic normalization, prior information, and planning assumptions differ.
arXiv:2610.14996v1 Announce Type: new Abstract: Choosing a few robust algorithms or schedules can require evaluating every candidate under many adverse scenarios, even though an agent only needs to i...
The paper studies the trade‑off between maximizing rewards and accurately estimating current treatment effects in nonstationary multi‑armed bandits. For $K$‑armed bandits with H"older‑smooth mean rewards, the authors derive minimax bounds on expected dynamic regret under a high‑probability simultaneous coverage requirement for all current effects. They propose the TRACE algorithm, which achieves a regret–precision trade‑off that separates reward‑learning regret from the unavoidable cost of refreshing current‑effect estimates, and extend the method with an empirical‑Bernstein e‑process to adapt exploration while preserving worst‑case guarantees.
BCPPO is a new variant of Proximal Policy Optimization that uses Bachelier-inspired cost‑prediction networks to generate a smooth penalty based on disagreement among critics. The method keeps temporal‑difference learning unchanged, applies a saturation‑aware controller to manage cost penalties, and deploys only the policy network. Across extensive experiments, BCPPO outperforms comparators in achieving higher mean returns while maintaining lower or comparable CVaR in all tested tasks.
arXiv:2606. 08791v1 Announce Type: cross Abstract: We study the problem of auditing a black-box algorithmic decision-maker from observable inputs and outputs alone.
The paper investigates best‑arm identification (BAI) in multi‑objective bandits, showing that additional objectives can actually reduce the number of samples needed. For Gaussian bandits, a characteristic‑time lower bound is derived that applies uniformly across preference orders, and explicit comparisons under Pareto, linear, and lexicographic preferences reveal that a larger optimal set can be identified with fewer samples, that the benefit of vector feedback depends on the preference, and that multi‑objective BAI can outperform the easiest single‑objective case. The authors also introduce Order‑Aware Track‑and‑Stop (OATS), an algorithm that achieves the lower bound while guaranteeing correctness at any desired confidence level.
The paper analyzes Bayesian linear bandits with isotropic Gaussian parameters, independent Gaussian arms, and Gaussian reward noise when the time horizon scales with the dimension. It derives explicit limits for the normalized posterior uncertainty and parameter overlaps, yielding exact regret curves for several policies—including Thompson sampling, posterior‑mean greedy selection, and scaled‑covariance variants. The results show that posterior‑mean greedy selection achieves the optimal Bayes regret, while Thompson sampling incurs a strictly larger leading regret whose ratio to greedy lies between one and two, approaching two for long horizons.