研究：用户自适应行为如何影响 LLM 交互系统的评估
When Both Sides Adapt: The Challenges of Evaluating Interactive LLM Systems
评估 LLM 时你可能一直测错了东西：这篇论文证明离线跑基准和线上 A/B 测的根本不是同一个指标，用户适应后模型排名都可能反转。做模型评估、部署选型或实验设计的工程师和研究者应该看看它给出的判断条件。
arXiv 论文（编号 2611.17239）研究了用户自适应行为与提示词内生性对 LLM 评估的影响：用户会根据模型响应改变后续提示策略和任务难度，因此模型表现会反过来影响被评估的提示分布。
论文定义了两类估计量：common-distribution effect 比较模型在同一提示分布上的表现，induced-distribution effect 比较各模型在用户交互实际产生的提示分布上的表现。作者证明离线基准测试与线上 A/B 测试分别收敛于这两类不同估计量，两者可能对模型给出相反的排名，并给出了排名反转的条件和保证一致的充分条件。
论文还在一个线性交互模型中证明用户自适应性会增大估计方差，并指出自适应性是 reward hacking 和新奇效应等评估问题的共同机制。结论是评估需同时测量模型性能与模型引发的用户行为变化。
When Both Sides Adapt: The Challenges of Evaluating Interactive LLM Systems
In interactive systems where users engage with LLMs, users are adaptive and, after observing LLM responses, can change their prompting strategy, as well as the type and difficulty of subsequent tasks. Thus model responses can influence the set of prompts on which the model is subsequently evaluated. We study the implications of this user adaptivity and prompt endogeneity for the evaluation of LLMs. We define two classes of estimands for comparing models: the common-distribution effect, which compares model performance on the same distribution of prompts, and the induced-distribution effect, which compares each model on the prompt distribution generated through user interaction. We show that offline benchmark evaluation and live A/B testing target different quantities, with the standard difference-in-means estimators for the two methods converging to the common-distribution effect and the induced-distribution effect, respectively. This distinction can affect decision-making: the two comparisons can rank models differently and lead to discrepancies in deployment recommendations. We characterize when such ranking reversals occur and provide a sufficient condition under which the rankings are guaranteed to agree. This result provides a practical heuristic for assessing whether offline evaluation is sufficient to reproduce the decision suggested by a live experiment. Third, we show, in a stylized linear model of interactions, that user adaptivity can increase variance in estimation. Finally, we discuss implications of user adaptivity for the interpretation of experiments and show that adaptivity creates a common mechanism leading to common evaluation challenges, such as reward hacking and novelty effects. Together, this work highlights the necessity of measuring both model performance as well as changes in the user behavior that the model induces.
本条情报归属 AITOP「论文」核心栏目，重点聚焦北美前沿实验室与硅谷科技巨头。在跨区域产业观察坐标中，该动态直接关联「AI 开发者工具链、Coding Agent 与工程化实践」研究主线，反映出当前技术演进与商业落地的最新风向。涉及核心实体与技术议题包括：LLM、A/B测试、模型评估、reward hacking，读者可通过下方标签进入对应专题页查阅历史脉络与关联对比。
信源核验说明：本文原始线索采编自「arXiv cs.AI」（@Hannah Li），经 AITOP 多维评分管线结构化提炼。