Large Language Models Fail on Trivial Alterations to Theory-of-Mind Tasks

TLDR

Intuitive psychology underpins common‑sense reasoning, and recent benchmarks for large language models focus on belief attribution in Theory‑of‑Mind tasks. The study tests whether large language models truly capture Theory‑of‑Mind reasoning by challenging a recent success case with subtle variations and considering the implications of future successes for human ToM tasks. The authors re‑examined the success case by applying small, principle‑preserving variations to the tasks, which reversed the results. The findings show that while some tasks succeed, large language models fail on trivial alterations, supporting a skeptical zero‑hypothesis and giving weight to outlier failures over average success.

Abstract

Intuitive psychology is a pillar of common-sense reasoning. The replication of this reasoning in machine intelligence is an important stepping-stone on the way to human-like artificial intelligence. Several recent tasks and benchmarks for examining this reasoning in Large-Large Models have focused in particular on belief attribution in Theory-of-Mind tasks. These tasks have shown both successes and failures. We consider in particular a recent purported success case, and show that small variations that maintain the principles of ToM turn the results on their head. We argue that in general, the zero-hypothesis for model evaluation in intuitive psychology should be skeptical, and that outlying failure cases should outweigh average success rates. We also consider what possible future successes on Theory-of-Mind tasks by more powerful LLMs would mean for ToM tasks with people.