Optimal Stopping Policies for Self-Refining Foundation Models

A new research paper on arXiv proposes an optimal stopping framework for self-refining foundation models, where models iteratively generate outputs, receive feedback from verifiers, and refine responses via in-context learning. The study formalizes the number of refinement iterations as an optimal stopping problem, balancing expected improvement against cost. The authors derive optimal stopping policies computable via stochastic approximation and test them on a coding benchmark, showing significant cost efficiency over prior stopping policies. The paper is categorized under Electrical Engineering and Systems Science, specifically Systems and Control, and was submitted to arXiv (ID: 2608.10729). This work addresses the growing need for efficient deployment of large language models in iterative refinement tasks, potentially impacting AI-driven art generation and curation tools.