Does Theory of Mind Improvement Really Benefit Human-AI Interactions? Empirical Findings from Interactive Evaluations

TL;DR

This paper evaluates whether improvements in Theory of Mind (ToM) capabilities of Large Language Models translate into better human-AI interactions, emphasizing the importance of interaction-based assessments over static benchmarks.

Contribution

It introduces a new interactive ToM evaluation paradigm and systematically studies four ToM enhancement techniques across real-world datasets and user interactions.

Findings

Static benchmark improvements do not always improve dynamic HAI performance.

Interaction-based assessments are crucial for developing socially aware LLMs.

The study covers both goal-oriented and experience-oriented tasks.

Abstract

Improving the Theory of Mind (ToM) capability of Large Language Models (LLMs) is crucial for effective social interactions between these AI models and humans. However, the existing benchmarks often measure ToM capability improvement through story-reading, multiple-choice questions from a third-person perspective, while ignoring the first-person, dynamic, and open-ended nature of human-AI (HAI) interactions. To directly examine how ToM improvement techniques benefit HAI interactions, we first proposed the new paradigm of interactive ToM evaluation with both perspective and metric shifts. Next, following the paradigm, we conducted a systematic study of four representative ToM enhancement techniques using both four real-world datasets and a user study, covering both goal-oriented tasks (e.g., coding, math) and experience-oriented tasks (e.g., counseling). Our findings reveal that improvements on static benchmarks do not always translate to better performance in dynamic HAI interactions. This paper offers critical insights into ToM evaluation, showing the necessity of interaction-based assessments in developing next-generation, socially aware LLMs for HAI symbiosis.