用真实用户数据评估大模型健康教练，发现统一多工具策略伤害特定群体。
Offline Policy Evaluation of Multi-Turn LLM Health Coaching with Real Users
- 分解决策头（工具/风格）实现离线评估，识别策略优劣
- 轻量模拟器加早期信息增益奖励，提升特质识别速度与目标达成率
- 建议先评估再个性化，按子群体制定指标避免平均值掩盖问题
我们研究了一个在网页上部署、具备工具增强功能的大型语言模型健康教练系统，基于7名真实用户的试点数据（280个评分回合）。通过因子化决策头（工具/风格）进行离线策略评估（OPE），发现采用统一重工具策略虽提升整体平均价值，却对低健康素养但高自我效能的用户群体造成负面影响。一个包含隐式原型的轻量级模拟器进一步显示，加入微小的早期信息增益奖励可稳定缩短特质识别时间，并提升目标成功率与pass@3指标。这些初步结果表明，应采取‘评估先行’的个性化路径：冻结生成器，基于类型化奖励（客观工具结果与满意度）学习子群感知的决策头，并始终报告各原型的指标，以揭示被平均值掩盖的子群伤害。
原文摘要 · Abstract (English)
We study a web-deployed, tool-augmented LLM health coach with real users. In a pilot with seven users (280 rated turns), offline policy evaluation (OPE) over factorized decision heads (Tool/Style) shows that a uniform heavy-tool policy raises average value on logs but harms specific subgroups, most notably low-health-literacy/high-self-efficacy users. A lightweight simulator with hidden archetypes further shows that adding a small early information-gain bonus reliably shortens trait identification and improves goal success and pass@3. Together, these early findings indicate an evaluation-first path to personalization: freeze the generator, learn subgroup-aware decision heads on typed rewards (objective tool outcomes and satisfaction), and always report per-archetype metrics to surface subgroup harms that averages obscure.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。