【主讲人简介】：Fengxiang He is a Lecturer (Assistant Professor) at the University of Edinburgh. His research is in (1) understanding AI behaviour, from both learning-theoretical and game-theoretical views; (2) developing models and algorithms that leverage the intrinsic symmetries in data, task, environment, etc.; (3) collaboration and interactions between AI agents in decentralised and multi-agent settings; and (4) employing AI to solve game-theoretical, economic, and financial problems. He is an Area Chair of ICML, NeurIPS, ICLR, UAI, AISTATS, and EMNLP (Industry Track), and on the Editorial Boards of Machine Learning and Pattern Recognition.

【内容简介】：LLMs, and other foundation models, are increasingly embedded in real-world, high-stakes systems that directly impact human lives and social fabric. This calls for a substantial understanding of its behaviours through a theoretical lens. In the first part of this talk, I will discuss the generalisability of reinforcement learning from human feedback (RLHF), an algorithm that LLM reasoning heavily relies on. Our recent work shows that the generalisation error of RLHF is largely determined by (1) a sampling error from sampling prompts and rollouts, (2) reward shift between the reward model, trained on preference data from earlier or mixed behaviour policies, and the current policy on rollouts, and (3) clipped KL regularisation that estimates KL regulariser from sampled log-probability ratios and then clip it for stabilisation. I will also discuss a few special cases and practical implications of the theory. In the second part, I will discuss the rationality of RL agents in deployment, which refers to the capability of optimising utilities. Rationality is fundamental in modelling game-theoretical behaviour of AI agents, yet it is considerably understudied in the literature. Our recent work proposes a suite of rationality measures to address this issue. We show that the rationality in deployment is determined by (1) an extrinsic component caused by environment shifts between training and deployment, and (2) an intrinsic one due to the algorithm's generalisability in a dynamic environment. I will also discuss the connections and differences between generalisability and rationality.

【讲座时间】：2026年8月4日（星期二）下午15:00

【讲座地点】：人文社科科研楼1801会议室