时间：2026年6月4日（周四）中午 12:30
地点：清华大学吕大龙楼11层1100
宛小昂课题组
报告主题 The Belief Conversion Asymmetry: A Cue-to-Belief Framework for Understanding Biased Reliance on AI Advisors
报告人：刘梦颖
内容：
AI advisory systems are increasingly deployed in high-stakes decision-making, yet how individuals convert AI competence information into calibrated reliance behavior remains poorly understood. Across three studies using a judge-advisor paradigm (N = 1,080), we examined how performance metrics drive advice-taking from AI versus human advisors at behavioral and metacognitive levels.
Study 1 revealed a framing-dependent asymmetry in competence label-sensitive calibration: accuracy labels calibrated both advice adoption and confidence change for human but not AI advisors, whereas error-rate labels eliminated competence label-sensitive calibration across both advisor types.
Study 2 demonstrated a framing-independent pattern in which internally held competence beliefs symmetrically calibrated both outcomes across AI and human advisors.
Study 3 found that participants tracked AI and human advisors’ performance with comparable sensitivity but formed less stable accuracy estimates for AI advisors. After controlling for estimate stability, newly acquired competence estimates produced symmetric effects across advisor types, with both predicting confidence change, though insufficient to predict advice adoption. This finding suggests that once competence beliefs are stabilized, AI input can be weighted equivalently to human input.
Collectively, our findings locate the AI-human asymmetry in competence-based calibration at the stage where external performance cues are converted into internal competence beliefs, not where those beliefs guide advice weighting. Effective human-AI collaboration thus requires a shift from merely making AI competence visible to helping users internalize AI performance information as stable and actionable competence beliefs.
郑美红课题组
报告主题 项目匹配还是结构化表征？工作记忆中的时长组织机制
报告人：赵云欣
内容：
多个时长在工作记忆中究竟如何被组织并用于后续判断，是时间认知领域尚未解决的重要问题，也关系到工作记忆究竟只是保存离散项目，还是能够从连续经验中抽取结构并形成内部组织这一更基础的问题。传统观点通常将时长记忆理解为一组独立的项目痕迹，即个体在再认时将探测时长与已编码的时长进行逐一匹配，并据此判断其是否曾经出现。然而，这一观点忽略了一种关键可能性：当多个时长在同一情境中被连续编码时，工作记忆保存的或许并不是彼此孤立的时长，而可能是围绕序列整体结构组织起来的时间信息。
本研究通过两个时长再认实验，结合计算建模，考察时长再认所依赖的多层级时间信息。结果显示，再认判断不仅受到局部项目相似性的影响，也受到任务水平全局范围和试次水平中心信息的共同塑造。尤其值得注意的是，未曾呈现但位于学习序列中心的探测时长能够诱发稳定的虚假再认，表明时长再认并不完全取决于某一时长是否真实出现，还会受到序列整体结构的系统性影响。进一步分析发现，基本时间精度主要解释个体对局部项目信息的利用，却不能可靠解释其对全局范围和序列中心信息的敏感性。在控制时间精度后，被试对序列中心信息和局部项目相似性的依赖仍呈负向关联，提示时长再认可能包含两种不同的加工取向：一种是面向摘要结构的整体加工，另一种是面向具体项目的局部匹配。
基于这些发现，本研究提出结构化时长组织观点。该观点认为，多个时长在工作记忆中并非只是独立痕迹的集合，而可能被组织为一种包含全局边界、序列中心和局部项目信息的多层级结构化表征。本研究挑战了将时长再认简化为项目匹配的传统理解，揭示了时间工作记忆在再认判断中并非只是被动保存和匹配离散时长，而可能从连续时间经验中抽取结构信息，并通过多层级表征读取来支持后续判断。本研究不仅拓展了对时长再认机制的理解，也为揭示工作记忆如何将连续经验转化为结构化内部模型提供了新证据。
