"Mind Your Theory: Theory of Mind Goes Deeper Than Reasoning"

LLMs need both correct mental state attribution and appropriate depth of mentalization for true Theory of Mind (ToM) .

Current ToM benchmarks miss half the puzzle: knowing when to mentalize matters as much as how.

This paper identifies a critical gap in evaluating Theory of Mind (ToM) capabilities in LLMs, highlighting the need to assess both depth of mentalization and correct inference.

-----

https://arxiv.org/abs/2412.13631

🤔 Original Problem:

→ Current AI research focuses solely on testing whether LLMs can correctly attribute mental states, ignoring whether ToM should be invoked in the first place.

→ Existing benchmarks use static scenarios that don't properly evaluate interactive ToM capabilities.

-----

🔬 Solution in this Paper:

→ The paper proposes a two-step evaluation framework for ToM capabilities.

→ First step determines whether to invoke ToM and at what depth of mentalization.

→ Second step applies correct inference given the chosen depth.

→ This approach distinguishes between three types of ToM errors: unnecessary ToM use, insufficient depth, and incorrect reasoning.

-----

🎯 Key Insights:

→ Biological agents adaptively choose ToM depth based on context and resource constraints

→ Linear probing methods cannot distinguish between different types of ToM failures

→ Interactive benchmarks are needed to test appropriate ToM invocation

-----

📊 Results:

→ Current benchmarks fail to distinguish between Type B (insufficient depth) and Type C (incorrect reasoning) errors

→ Static vignette-based tasks preclude testing true ToM capabilities