时间：2026年5月28日（周四）中午 12:30
地点：清华大学吕大龙楼11层1100
报告主题：Emotion Utilisation in Social Norm Enforcement by Large Language Models
报告人：谈昊天
内容：
Understanding how large language models (LLMs) utilise emotion-like representations to enforce social norms is crucial for LLM deployment in real-world, yet it is rarely explored. We investigated this in altruistic third-party punishment, comparing 61,020 decisions from 1,017 humans with over 2.98 million decisions from 21 LLMs. Results showed that LLMs punished unfairness more often and showed divergent information processing and emotion utilisation compared to humans. Humans exhibit gradual emotional and behavioral changes in response to contextual information, and their social norm enforcement relies on emotional valence but not arousal. In contrast, LLMs exhibited threshold-like decision geometry, in which decisions shifted abruptly at fair-to-unfair allocations and punishment-cost midpoints; and they amplified emotion–behavior coupling regardless of valence or arousal. Besides, later-released models with higher capability showed greater alignment with humans, suggesting that socio-emotional capabilities may co-evolve with cognitive capability. Causally, sparse autoencoding identified a specific neuron such that intervention on it simultaneously regulated emotion-like representations and norm enforcement. These findings identify emotion-like representations as internal signals in LLMs' norm enforcement and reveal non-human threshold-like and arousal-conflated dynamics, providing implications for the regulation and deployment of LLMs.
