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

TL;DR

This paper demonstrates that large language models often fail on simple variations of Theory-of-Mind tasks, questioning their genuine understanding of intuitive psychology and common-sense reasoning.

Contribution

The study critically evaluates recent claims of success in LLMs on ToM tasks, highlighting their fragility to trivial task alterations and emphasizing the need for skeptical evaluation.

Findings

LLMs fail on trivial variations of ToM tasks

Success rates are unreliable without considering failure cases

Zero-hypothesis should be skeptical of LLMs' ToM abilities

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.