A Conceptual Primitive Decomposition of the Sally-Anne Test

Document Type

Conference Proceeding

Publication Date

Fall 2025

Publication Title

Proceedings of the 2025 AAAI Fall Symposium Series

Abstract

Although large language models (LLMs) have been observed to perform at a human level in theory of mind tasks, deeper examinations and systematic testing of their performance in these domains is needed. Primitive decomposition representations show promise for building robotic systems with greater abilities for in-depth natural language understanding and generation. In this work, we explore representations of theory of mind which are combinations of conceptual primitives, focusing on simulations of a Sally-Anne false-belief test. We demonstrate how primitive decompositions into the conceptual building blocks of image schemas and Conceptual De- pendency can represent the attribution of false beliefs to intelligent agents. The exploration has consequences for generating controlled and linguistically varied tests posed in natural language as challenge problems for large language models and for cognitive representations more broadly.

Volume

7

Issue

1

First Page

670

Last Page

672

ISSN

2994-4317

Comments

The Association for the Advancement of Artificial Intelligence’s 2025 Fall Symposium Series (FSS-25) was held on November 6-8, 2025 in Arlington, Virginia.

Link to Resource

Share

COinS