Isolating Functional ToM Reasoning
Background: Theory of Mind (ToM) benchmarks for LLMs often struggle to distinguish between genuine functional reasoning and the exploitation of lexical shortcuts or spurious correlations in conversation history.
Question / Future Work: It remains an open challenge to develop benchmarks that effectively isolate functional ToM from surface-level linguistic pattern matching, specifically by testing model performance on state-consistent trajectory prediction in the absence of informative context.
Why It Matters: This distinguishes between predictive linguistic pattern matching and actual agentic reasoning, a requirement for reliable social intelligence in future models.
Metadata & Links
- created_at
- 2026-04-25T05:37:42Z