Reproduce “Towards Safety Evaluations of Theory of Mind in Large Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Tatsuhiro Aoshima, Mitsuaki Akiyama
A research team from NTT developed an evaluation framework to assess whether concerning behaviors in large language models stem from sophisticated "scheming reasoning" or complex pattern matching, using Theory of Mind tasks. Their evaluations of open-weight LLMs showed consistently low performance on false belief tasks, indicating an absence of the cognitive capabilities for intentional deception.
Key insights
- The observed hierarchy of task difficulty in LLMs (True Belief < Mental State Attribution < False Belief) mirrors the developmental trajectory of ToM in humans.
- ToM capabilities, particularly for advanced False Belief tasks, show inconsistent 'developmental' trends across LLM versions and series, with improvements in one area not guaranteeing overall ToM advancement.
- Post-training (e.g., instruct-tuning) consistently improves general language comprehension but can lead to an average decline in performance on complex False Belief tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.