How FaR Are Large Language Models From Agents with Theory-of-Mind?
Pei Zhou♢♢{}^{\diamondsuit}start_FLOATSUPERSCRIPT ♢ end_FLOATSUPERSCRIPT Aman Madaan♠♠{}^{\spadesuit}start_FLOATSUPERSCRIPT ♠ end_FLOATSUPERSCRIPT Srividya Pranavi Potharaju ††{}^{\dagger}start_FLOATSUPERSCRIPT † end_FLOATSUPERSCRIPT Aditya Gupta Kevin R. McKee‡normal-‡{}^{\ddagger}start_FLOATSUPERSCRIPT ‡ end_FLOATSUPERSCRIPT Ari Holtzman♣normal-♣{}^{\clubsuit}start_FLOATSUPERSCRIPT ♣ end_FLOATSUPERSCRIPT Jay Pujara♢normal-♢{}^{\diamondsuit}start_FLOATSUPERSCRIPT ♢ end_FLOATSUPERSCRIPT Xiang Ren♢normal-♢{}^{\diamondsuit}start_FLOATSUPERSCRIPT ♢ end_FLOATSUPERSCRIPT Swaroop Mishra‡normal-‡{}^{\ddagger}start_FLOATSUPERSCRIPT ‡ end_FLOATSUPERSCRIPT Aida Nematzadeh‡normal-‡{}^{\ddagger}start_FLOATSUPERSCRIPT ‡ end_FLOATSUPERSCRIPT Shyam Upadhyay†normal-†{}^{\dagger}start_FLOATSUPERSCRIPT † end_FLOATSUPERSCRIPT Manaal Faruqui†normal-†{}^{\dagger}start_FLOATSUPERSCRIPT † end_FLOATSUPERSCRIPT ††\dagger† Google ‡‡\ddagger‡ Google DeepMind ♢♢\diamondsuit♢ University of Southern California ♠♠\spadesuit♠ Carnegie Mellon University ♣♣\clubsuit♣ University of Chicago Work done during Google Internship
Abstract
Thinking is for Doing. Humans can infer other people’s mental states from observations–an ability called Theory-of-Mind (ToM)–and subsequently act pragmatically on those inferences. Existing question answering benchmarks such as ToMi ask models questions to make inferences about beliefs of characters in a story, but do not test whether models can then use these inferences to guide their actions. We propose a new evaluation paradigm for large language models (LLMs): Thinking for Doing (T4D), which requires models to connect inferences about others’ mental states to actions in social scenarios. Experiments on T4D demonstrate that LLMs such as GPT-4 and PaLM 2 seemingly excel at tracking characters’ beliefs in stories, but they struggle to translate this capability into strategic action.
中文速览
大语言模型（LLM）在"心智理论"（Theory of Mind, ToM）测试中能答对"角色会去哪里找东西"这类推断题，却往往不知道该据此采取什么行动——这正是本文要解决的问题。作者提出了一个新评测框架"边想边做"（Thinking for Doing, T4D），要求模型像真实智能体一样，根据对他人信念的推断来选择最恰当的助人行动，而非仅仅回答关于心理状态的问题。实验发现，GPT-4在T4D上的零样本准确率只有50%，而人类标注者的一致率超过95%，核心瓶颈在于模型难以自发识别出隐含的心智状态推断步骤。为此，作者提出了零样本提示框架"预见与反思"（Foresee and Reflect, FaR）：先让模型预判角色将面临的困境，再反思哪种行动最能帮到对方，从而将GPT-4的表现从50%提升至71%，并在分布外场景中持续优于思维链（Chain-of-Thought）等其他提示方法。这项工作揭示了当前LLM从"能推断"到"会行动"之间的关键鸿沟，对开发真正具备社会智能的AI智能体具有重要意义。
原文 arXiv:2310.03051;中英对照 + 大白话阅读 https://aha.fim.ai/paper/2310.03051v1