报告题目:Social intelligence: modeling, evaluation and challenges of LLMs
报告人:王俊淇 北京通用人工智能研究院研究员
报告摘要:
In cognition point of view, human intelligence is a set of cognitive abilities that enable individuals to perceive, understand, and interact with the world around them. Social intelligence, as a subset of human intelligence, refers to the ability to understand and navigate social relationships, emotions, and behaviors. In this talk, we start from Theory of Mind (ToM), a quite fundamental element in social intelligence. We introduce a mathematical framework of ToM in cooperative communication point of view and see how modern LLMs bring new challenges. Then I will talk about the our evaluations on social intelligence on LLMs. At last, I will talk about our new directions in research, these new thinkings and explorations are considered as our reactions to the challenges of LLMs.
报告人简介:
王俊淇博士,长期致力于对智能现象、认知行为及其计算机制的探索,研究内容涵盖对社会智能、语言模型的机制解释和相应的测试方法,以及人类学习过程的建模与神经机制探索。在ICML,NeurIPS,ICLR,CogSci,Trends in Cognitive Science,Entorpy等会议及期刊上发表文章近20篇。在社会智能方面,提出了Generalized Belief Transport理论框架,解释智能体合作条件下的数学机理,以心智理论为基础初步揭示并统一了各种合作行为模式,并对其迭代动力学进行了数学分析,其中统一模型的工作在NeurIPS会议上做了Oral报告。同时,以语言模型同时作为研究对象和研究工具,提出了心智理论方面的测试方法,以及基于《文明》游戏、复杂交易系统TradeCraft的复杂学习推理环境和测试基准。此外,在人类价值建模和维度结构方面也做出了相关工作。
报告时间:2025年8月25日(星期一)8:30-9:30
报告地点:吉林大学正新楼三楼报告厅
主办单位:吉林大学人工智能学院
