FANToM: A Benchmark for Stress-testing Machine Theory of Mind in Interactions
Hyunwoo Kim, Melanie Sclar, Xuhui Zhou, Ronan Le Bras, Gunhee Kim, Yejin Choi, Maarten Sap
TL;DR
FANToM introduces an interactive, information-asymmetric benchmark to stress-test theory of mind (ToM) reasoning in conversations, addressing biases inherent in narrative-based ToM tasks. It establishes a six-type ToM QA framework derived from FactQ, spanning belief attribution and access awareness, and evaluates models using short and full conversation contexts with diverse question formats. Empirical results show current LLMs lag behind humans in ToM tasks, with chain-of-thought prompting offering modest gains and fine-tuning sometimes surpassing human performance on isolated types but not on coherent All-metrics. The work underscores the importance of interaction-centric evaluation to reveal illusory ToM and points to future directions such as grounding reasoning and exploring multi-modal or more dynamic social settings.
Abstract
Theory of mind (ToM) evaluations currently focus on testing models using passive narratives that inherently lack interactivity. We introduce FANToM, a new benchmark designed to stress-test ToM within information-asymmetric conversational contexts via question answering. Our benchmark draws upon important theoretical requisites from psychology and necessary empirical considerations when evaluating large language models (LLMs). In particular, we formulate multiple types of questions that demand the same underlying reasoning to identify illusory or false sense of ToM capabilities in LLMs. We show that FANToM is challenging for state-of-the-art LLMs, which perform significantly worse than humans even with chain-of-thought reasoning or fine-tuning.