A new arXiv preprint studies fixed-budget adaptive experimentation, where samples are allocated across designs and a single design is recommended at the end. The paper's goal is not simply to find the best arm: it aims to maximize the probability that the recommended design's mean performance exceeds a target with sufficient posterior evidence. The authors call this objective posterior success.

The proposed method, Halve-Then-Greedy, is presented as an adaptive sampling rule for this setting. The abstract emphasizes that the algorithm achieves near-optimal instance dependence, meaning its performance adapts to the difficulty of the particular problem rather than relying on a worst-case guarantee. The approach is explicitly framed in contrast to best-arm identification, though the provided abstract cuts off before giving the full distinction.

As stated in the abstract, the paper's contribution is a sampling strategy for settings where a decision is judged by whether a recommended design clears a threshold, rather than by how well it ranks among alternatives. The abstract does not detail the halving or greedy steps in the text available, but the headline claim is improved posterior success with near-optimal instance dependence.