AV-AIVAT：基于不完全信息博弈中认证随时有效停止机制，将智能体评估成本降低 74 倍

文章背景与核心概要

在扑克等不完全信息博弈中，评估和比较两个智能体的性能是一项极其消耗计算资源的任务，需要大量的对局来克服运气带来的方差。传统的固定预算评估方法往往会在胜负已分后继续白白消耗资源，或者在能够可靠地区分智能体之前过早终止。同时，简单的连续测试方法又会违反统计置信度保证。

本文介绍了 AV-AIVAT（随时有效且结合动作信息的价值评估工具，Anytime-Valid Action-Informed Value Assessment Tool）框架，该框架将方差缩减技术与持续监控的置信序列（CSs）相结合。通过确保在线价值模型仅从历史对局中学习，AV-AIVAT 提供了经认证、可审计的提前停止机制——在严格保持统计有效性的同时，将评估成本降低了高达 \(74\times\)。

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

Authors: Boning Li, Yu Chen, Longbo Huang
Subjects: Computer Science and Game Theory (cs.GT); Artificial Intelligence (cs.AI); Computation and Language (cs.CL); Machine Learning (cs.LG); Multiagent Systems (cs.MA)
ArXiv ID: arXiv:2608.06362
Submitted: August 6, 2026

作者： Boning Li, Yu Chen, Longbo Huang
学科： 计算机科学与博弈论 (cs.GT)；人工智能 (cs.AI)；计算与语言 (cs.CL)；机器学习 (cs.LG)；多智能体系统 (cs.MA)
ArXiv ID： arXiv:2608.06362
提交时间： 2026年8月6日

Executive Summary

Evaluating and comparing the performance of two agents in imperfect-information games (such as poker) is resource-intensive, requiring extensive game-play to overcome variance from luck. Traditional fixed-budget evaluations either overspend after a clear winner emerges or terminate prematurely before agents can be reliably distinguished. Meanwhile, naive sequential testing violates statistical confidence guarantees.

This paper introduces AV-AIVAT (Anytime-Valid Action-Informed Value Assessment Tool), a framework that combines variance-reduction techniques with continuously monitored Confidence Sequences (CSs). By ensuring online value models learn exclusively from past games, AV-AIVAT delivers certified, auditable early stopping—reducing evaluation costs by up to \(74\times\) while strictly preserving statistical validity.

执行摘要

在扑克等不完全信息博弈中，评估和比较两个智能体的性能是一项资源密集型任务，需要进行大量的对局以克服运气带来的方差。传统的固定预算评估方法要么在明确的赢家出现后过度消耗资源，要么在智能体能够被可靠区分之前过早终止。同时，朴素的序贯测试会违反统计置信度保证。

本文引入了 AV-AIVAT（Anytime-Valid Action-Informed Value Assessment Tool，随时有效且结合动作信息的价值评估工具）框架，该框架将方差缩减技术与持续监控的置信序列（CSs）相结合。通过确保在线价值模型仅从过去的对局中学习，AV-AIVAT 提供了经认证、可审计的提前停止机制——在严格保持统计有效性的同时，将评估成本降低了高达 \(74\times\)。

Key Contributions & Methodology

The Stopping Problem in Agent Evaluation: Evaluating game-playing agents requires balancing accuracy and resource consumption (money, model inference, and expert time). Standard fixed-budget or heuristic stopping approaches either waste computational resources or invalidate statistical significance levels.
Integrating AIVAT with Confidence Sequences:
The baseline Action-Informed Value Assessment Tool (AIVAT) applies conditional mean-zero corrections to drastically lower variance in imperfect-information games (achieving a median \(54\times\) reduction across 15 LLM agent configurations over 71,439 paired hands of Heads-Up No-Limit Hold'em), but lacks a principled stopping rule.
AV-AIVAT pairs AIVAT with Confidence Sequences (CSs) to allow continuous monitoring without incurring multiple-testing errors.
Online Value Modeling: The underlying value model learns strictly from historical games, guaranteeing that no game is scored by its own correction.
Dual Stopping Framework:
Asymptotic Screening (AsympCS): At a nominal 95% level and a target precision of \(\pm1\) Big Blind, raw outcomes require a median \(74\times\) as many hands to stop compared to AIVAT-corrected outcomes.
Exact Finite-Sample Certification (EB-CS): Utilizes the Empirical-Bernstein CS, which relies on a structurally established bound for corrected payoffs (demonstrated on Leduc hold'em). Descriptive HUNL EB-CS evaluations demonstrate a median \(1.37\times\) stopping-time ratio.

核心贡献与方法论

智能体评估中的停止问题： 评估博弈智能体需要在准确性与资源消耗（金钱、模型推理和专家时间）之间取得平衡。标准的固定预算或启发式停止方法要么浪费计算资源，要么使统计显著性水平失效。

将 AIVAT 与置信序列相结合：

基线结合动作信息的价值评估工具（AIVAT）应用条件均值为零的修正，显著降低了不完全信息博弈中的方差（在对局总数达 71,439 手的单挑无限注德州扑克（HUNL）中，对 15 种大语言模型智能体配置实现了中位数 \(54\times\) 的方差缩减），但它缺乏一个原则性的停止规则。

AV-AIVAT 将 AIVAT 与置信序列（CSs）配对，允许进行持续监控而不会引发多重检验错误。

在线价值建模： 底层价值模型严格从历史对局中学习，从而保证没有任何对局是由其自身的修正来进行评分的。

双重停止框架：

渐近筛查（AsympCS）： 在名义 95% 置信水平和 \(\pm1\) 大盲注的目标精度下，与 AIVAT 修正后的结果相比，原始结果需要中位数达到 \(74\times\) 的对局手数才能停止。

精确有限样本认证（EB-CS）： 利用经验伯恩斯坦置信序列（Empirical-Bernstein CS），该序列依赖于修正后收益的结构性确立界限（在 Leduc 扑克上进行了演示）。对 HUNL EB-CS 的描述性评估表明，其停止时间比的中位数为 \(1.37\times\)。

Conclusion & Impact

AV-AIVAT bridges the gap between advanced variance reduction and sequential statistical testing. By separating asymptotic screening from exact finite-sample certification, the framework allows evaluations to halt the exact moment sufficient evidence is gathered. This provides third parties with a fully auditable trail to independently verify the verdict at the precise stopping timestamp.

结论与影响

AV-AIVAT 弥合了高级方差缩减与序贯统计测试之间的鸿沟。通过将渐近筛查与精确的有限样本认证分离开来，该框架允许评估在收集到足够证据的精确瞬间停止。这为第三方提供了一条完全可审计的追踪路径，以便在确切的停止时间戳独立验证评估结论。