Reproduce “Language Models Represent Beliefs of Self and Others” 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
Wentao Zhu, Zhining Zhang, Yizhou Wang
Peking University researchers demonstrated that large language models develop linearly decodable internal representations of their own and other agents' beliefs. Manipulating these internal representations causally influences the models' Theory of Mind capabilities, specifically improving performance on challenging false-belief tasks.
Key insights
- Large language models internally track distinct belief states for self and other agents, including those that diverge from factual reality.
- These internal belief representations are not merely correlates but play a causal role in the model's social reasoning processes, particularly in false-belief understanding.
- The identified belief representations and intervention directions generalize across different types of Theory of Mind tasks and external datasets, indicating a robust underlying mechanism.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.