Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

All You Need is Sally-Anne: ToM in AI Strongly Supported After Surpassing Tests for 3-Year-Olds

Nitay Alon, Joseph Barnby, Reuth Mirsky, Stefan Sarkadi

TL;DR

The paper investigates AI Theory of Mind (ToM) by introducing GeRRI, a framework that combines Gradient-Based Inference with Recursive Representation and Representation Learning. It grounds belief updates in Bayesian form $P(H|D) = \frac{P(D|H) P(H)}{P(D)}$ and optimizes using $\mathcal{L} = \mathrm{KL}(P(H|D) || Q(H))$ with entropy regularization, integrated via adaptive SGD and RNN-based recursion. Experimental results on Sally-Anne and Smarties tasks show the model achieves performance comparable to 3-year-old children across 100 trials per task, suggesting ToM-like reasoning is attainable in AI under gradient-based, recursive schemes. The authors also critique current ToM benchmarks and caution about the limitations of gradient-based hierarchical belief modeling, calling for broader, dynamic social scenarios and alternative frameworks within the ToM4AI initiative.

Abstract

Theory of Mind (ToM) is a hallmark of human cognition, allowing individuals to reason about others' beliefs and intentions. Engineers behind recent advances in Artificial Intelligence (AI) have claimed to demonstrate comparable capabilities. This paper presents a model that surpasses traditional ToM tests designed for 3-year-old children, providing strong support for the presence of ToM in AI systems.

All You Need is Sally-Anne: ToM in AI Strongly Supported After Surpassing Tests for 3-Year-Olds

TL;DR

The paper investigates AI Theory of Mind (ToM) by introducing GeRRI, a framework that combines Gradient-Based Inference with Recursive Representation and Representation Learning. It grounds belief updates in Bayesian form and optimizes using with entropy regularization, integrated via adaptive SGD and RNN-based recursion. Experimental results on Sally-Anne and Smarties tasks show the model achieves performance comparable to 3-year-old children across 100 trials per task, suggesting ToM-like reasoning is attainable in AI under gradient-based, recursive schemes. The authors also critique current ToM benchmarks and caution about the limitations of gradient-based hierarchical belief modeling, calling for broader, dynamic social scenarios and alternative frameworks within the ToM4AI initiative.

Abstract

Theory of Mind (ToM) is a hallmark of human cognition, allowing individuals to reason about others' beliefs and intentions. Engineers behind recent advances in Artificial Intelligence (AI) have claimed to demonstrate comparable capabilities. This paper presents a model that surpasses traditional ToM tests designed for 3-year-old children, providing strong support for the presence of ToM in AI systems.

Paper Structure

This paper contains 12 sections, 3 equations, 2 figures.

Table of Contents

  1. Introduction
  2. Related Work
  3. ToM and GeRRI
  4. Gradient-Based Inference:
  5. Recursive Representation:
  6. Representation Learning:
  7. Experimental Results
  8. Sally-Anne Task:
  9. Smarties Task:
  10. Experimental Setup and Procedure:
  11. Results:
  12. Conclusion

Figures (2)

  • Figure 1: An illustration of a Theory of Mind process between two agents
  • Figure 2: A can of Smarties full of pencils.