Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Rethinking Theory of Mind Benchmarks for LLMs: Towards A User-Centered Perspective

Qiaosi Wang, Xuhui Zhou, Maarten Sap, Jodi Forlizzi, Hong Shen

TL;DR

This paper tackles the problem that current ToM benchmarks for LLMs repurpose human ToM tasks, which suffer theoretical, methodological, and evaluation flaws. It synthesizes psychology and AI literature to diagnose these limitations and argues for a user-centered, interaction-focused redefinition of ToM benchmarks grounded in human-AI interaction needs. Key contributions include a critique of single-dimension ToM assessment, emphasis on psychometric validity, and a shift toward dynamic, interactive, and multimodal evaluation frameworks like SOTOPIA and MUMA. The proposed framework aims to yield benchmarks that better predict real-world user experiences and support safe, effective human-AI collaboration across diverse contexts.

Abstract

The last couple of years have witnessed emerging research that appropriates Theory-of-Mind (ToM) tasks designed for humans to benchmark LLM's ToM capabilities as an indication of LLM's social intelligence. However, this approach has a number of limitations. Drawing on existing psychology and AI literature, we summarize the theoretical, methodological, and evaluation limitations by pointing out that certain issues are inherently present in the original ToM tasks used to evaluate human's ToM, which continues to persist and exacerbated when appropriated to benchmark LLM's ToM. Taking a human-computer interaction (HCI) perspective, these limitations prompt us to rethink the definition and criteria of ToM in ToM benchmarks in a more dynamic, interactional approach that accounts for user preferences, needs, and experiences with LLMs in such evaluations. We conclude by outlining potential opportunities and challenges towards this direction.

Rethinking Theory of Mind Benchmarks for LLMs: Towards A User-Centered Perspective

TL;DR

This paper tackles the problem that current ToM benchmarks for LLMs repurpose human ToM tasks, which suffer theoretical, methodological, and evaluation flaws. It synthesizes psychology and AI literature to diagnose these limitations and argues for a user-centered, interaction-focused redefinition of ToM benchmarks grounded in human-AI interaction needs. Key contributions include a critique of single-dimension ToM assessment, emphasis on psychometric validity, and a shift toward dynamic, interactive, and multimodal evaluation frameworks like SOTOPIA and MUMA. The proposed framework aims to yield benchmarks that better predict real-world user experiences and support safe, effective human-AI collaboration across diverse contexts.

Abstract

The last couple of years have witnessed emerging research that appropriates Theory-of-Mind (ToM) tasks designed for humans to benchmark LLM's ToM capabilities as an indication of LLM's social intelligence. However, this approach has a number of limitations. Drawing on existing psychology and AI literature, we summarize the theoretical, methodological, and evaluation limitations by pointing out that certain issues are inherently present in the original ToM tasks used to evaluate human's ToM, which continues to persist and exacerbated when appropriated to benchmark LLM's ToM. Taking a human-computer interaction (HCI) perspective, these limitations prompt us to rethink the definition and criteria of ToM in ToM benchmarks in a more dynamic, interactional approach that accounts for user preferences, needs, and experiences with LLMs in such evaluations. We conclude by outlining potential opportunities and challenges towards this direction.

Paper Structure

This paper contains 12 sections, 1 figure.

Table of Contents

  1. Introduction
  2. Evaluating Theory of Mind in Humans and Large Language Models
  3. Assessing Human ToM through ToM Tasks
  4. Current State of ToM Benchmarks for LLMs
  5. The Limitations of Appropriating ToM Tasks as LLM Benchmarks
  6. Theoretical Limitation
  7. Methodological Limitation
  8. Evaluation Limitation
  9. Towards User-Centered Theory of Mind Benchmark for LLMs
  10. Defining LLM ToM From a User-Centered Perspective
  11. Benchmarking LLM ToM in Dynamic and Interactional Social Contexts
  12. Conclusion

Figures (1)

  • Figure 1: An illustration of the Sally-Anne test commonly used to evaluate children's Theory of Mind. Figure reproduced from scassellati2001foundations [2001].