CUE：面向多轮基准测试的校准用户嵌入
原标题：CUEing User Simulators: Calibrated User Embeddings for Multi-Turn Benchmarking
AI 摘要
该论文提出 Calibrated User Embeddings (CUE) 框架，用于多轮交互基准测试中的用户模拟器校准。研究发现现有模拟器虽在表面风格上接近人类，但缺乏结果校准，即无法复现真实用户与同一智能体交互时的成功率和失败模式。CUE 通过编码真实会话并采样连续表示，再解码为 persona 命令来引导 LLM 模拟用户，无需训练即可在多个任务上更忠实地复现真实用户的智能体失败模式和成功率。
正文节选
highlightbox CUEing User Simulators Abstract Recent benchmarks rely on user simulators to evaluate AI agents in multi-turn interaction. While existing simulation techniques demonstrate surface fidelity to human style and behavior, ecologically valid interactive benchmarking also requires alignment in when and how agents fail across simulated and real user populations. We find that existing simulators lack outcome calibration: agreement with observed success rates and failure patterns when real u