Reproduce “TimeToM: Temporal Space is the Key to Unlocking the Door of Large Language Models' Theory-of-Mind” 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
Guiyang Hou, Wenqi Zhang, Yongliang Shen, Linjuan Wu, Weiming Lu
Zhejiang University researchers developed TIMETOM, a zero-shot prompting framework that enhances Large Language Models' (LLMs) Theory of Mind (ToM) capabilities through explicit temporal space modeling and a novel belief solver. The method yielded substantial accuracy gains across various benchmarks, particularly on challenging higher-order ToM questions, improving GPT-4's performance by 29.00% on second-order ToM tasks on ToMI.
Key insights
- Explicitly modeling the temporal dimension of events and how character beliefs evolve over time is fundamental for LLMs to achieve robust and coherent Theory of Mind reasoning.
- Higher-order Theory of Mind questions can be effectively addressed by identifying periods of shared information (belief communication) and transforming the complex inference into a temporally-constrained first-order problem.
- Structuring character beliefs into 'self-world' (direct knowledge) and 'social-world' (knowledge about others' beliefs) provides a cognitively inspired framework that helps LLMs differentiate and process various orders of ToM.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.