AV-AIVAT: 74× Cheaper Agent Evaluation with Certified Anytime-Valid Stopping in Imperfect-Information Games

Evaluating agents in imperfect-information games (IIGs), such as poker, is notoriously expensive and noisy. Traditional fixed-budget evaluations often waste resources by continuing after a winner is clear or fail to reach a conclusion if the budget is too small. While researchers use variance-reduction techniques like AIVAT to lower noise, these methods do not inherently provide a valid stopping rule. This paper addresses how to stop an evaluation as soon as sufficient evidence is gathered while maintaining rigorous statistical guarantees.

The authors introduce AV-AIVAT, a framework that integrates the Action-Informed Value Assessment Tool (AIVAT) with continuously monitored Confidence Sequences (CSs). The key innovation is a predictable interface that allows the value model to learn from past games without biasing the current game's correction. By ensuring that corrections are conditionally mean-zero and independent of the current hand's outcome, the authors preserve the validity of the statistical inference. The framework provides two types of confidence intervals: an asymptotic CS (AsympCS) for efficient, practical monitoring, and an Empirical-Bernstein CS (EB-CS) for exact, finite-sample certification.

AV-AIVAT significantly improves sample efficiency. Across 15 LLM agent configurations and over 71,000 hands of Heads-Up No-Limit Hold'em, the framework reduced the median number of hands required to reach a target precision of ±1 Big Blind by 74x using the AsympCS. The authors also establish a release protocol that allows third parties to audit and reconstruct the stopping claim using only the corrected payoff stream and metadata, ensuring that early-stopping decisions are transparent and reproducible.

This work bridges the gap between variance-reduction techniques and sequential statistical inference. It provides a principled way to conduct "anytime-valid" evaluations, allowing researchers to stop experiments the moment a verdict is statistically supported. This reduces the high costs associated with LLM inference and expert time in competitive game research, while simultaneously providing a framework for auditable, verifiable results.