时 间：2026年10月13日（周二） 16：00 - 17：00
地 点：中北理科大楼A1514室
报告人：张智恒 上海财经大学常任轨助理教授
主持人：章迎莹 华东师范大学统计学院
摘 要：
This talk addresses a fundamental question at the intersection of statistics and foundation models: can a foundation model learn not only what a (causal) effect is, but also how a statistical estimator should fluctuate under repeated sampling? This question has direct implications for statistical inference and uncertainty quantification. To address it, we propose Fluctuation-Supervised Pretraining (FSP), which jointly uses the average treatment effect and the sample-level fluctuation characterized by the efficient influence function as supervision during synthetic pretraining, thereby teaching the model how an efficient estimator should respond to a particular random sample. Theoretically, we show that full fluctuation is a statistically distinguished endpoint: any fixed partial fluctuation retains first-order label ambiguity, whereas full fluctuation improves label learnability to a second-order scale. We further characterize the role of finite-pretraining error. This framework connects classical semiparametric efficiency theory and influence functions with modern foundation-model pretraining, supporting the feasibility of a new training paradigm: rather than teaching a model only the correct answer, we seek to teach it the statistical principles that reliable answers should obey.
报告人简介：
张智恒现任上海财经大学统计与数据科学学院常任轨助理教授。此前，他于2025年8月在清华大学交叉信息研究院（姚班）获得博士学位。其研究主要聚焦于因果推断、实验设计与可信机器学习，关注在现实约束下，如何建立可识别、统计高效且可执行的因果学习方法。近期，他进一步关注如何将统计推断原理融入基础模型预训练，使传统上需要针对每个数据集重新执行的推断过程逐步转化为可预训练、可复用和可验证的智能推断和不确定性量化能力，并探索连接因果推断、实验设计、在线决策与可信人工智能的统一理论框架。张智恒担任2025年度 CCF—滴滴“盖亚”学术基金联合项目负责人，并担任 AAAI 2025 Artificial Intelligence with Causal Techniques（AICT）领域主席等。他的相关研究成果已发表于或在修于 JASA、JMLR、IJOO、COLT、ICML、NeurIPS、UAI、KDD、SIGIR 等期刊和会议。