Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ToM-LM: Delegating Theory of Mind Reasoning to External Symbolic Executors in Large Language Models

Weizhi Tang, Vaishak Belle

TL;DR

This study proposes a novel view about externalizing a particular component of ToM reasoning, mainly reasoning about beliefs, and suggests generalizing it to other aspects of ToM reasoning.

Abstract

Theory of Mind (ToM) refers to the ability of individuals to attribute mental states to others. While Large Language Models (LLMs) have shown some promise with ToM ability, they still struggle with complex ToM reasoning. Our approach leverages an external symbolic executor, specifically the SMCDEL model checker, and fine-tuning to improve the ToM reasoning ability of LLMs. In our approach, an LLM is first fine-tuned through pairs of natural language and symbolic formulation representation of ToM problems and is then instructed to generate the symbolic formulation with a one-shot in-context example. The generated symbolic formulation is then executed by the SMCDEL model checker to perform transparent and verifiable ToM reasoning and give the final result. We demonstrate that our approach, ToM-LM, shows a significant improvement over all the constructed baselines. Our study proposes a novel view about externalizing a particular component of ToM reasoning, mainly reasoning about beliefs, and suggests generalizing it to other aspects of ToM reasoning.

ToM-LM: Delegating Theory of Mind Reasoning to External Symbolic Executors in Large Language Models

TL;DR

This study proposes a novel view about externalizing a particular component of ToM reasoning, mainly reasoning about beliefs, and suggests generalizing it to other aspects of ToM reasoning.

Abstract

Theory of Mind (ToM) refers to the ability of individuals to attribute mental states to others. While Large Language Models (LLMs) have shown some promise with ToM ability, they still struggle with complex ToM reasoning. Our approach leverages an external symbolic executor, specifically the SMCDEL model checker, and fine-tuning to improve the ToM reasoning ability of LLMs. In our approach, an LLM is first fine-tuned through pairs of natural language and symbolic formulation representation of ToM problems and is then instructed to generate the symbolic formulation with a one-shot in-context example. The generated symbolic formulation is then executed by the SMCDEL model checker to perform transparent and verifiable ToM reasoning and give the final result. We demonstrate that our approach, ToM-LM, shows a significant improvement over all the constructed baselines. Our study proposes a novel view about externalizing a particular component of ToM reasoning, mainly reasoning about beliefs, and suggests generalizing it to other aspects of ToM reasoning.
Paper Structure (32 sections, 7 figures, 1 table)

This paper contains 32 sections, 7 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. ToM In LLMs
  4. Delegating to External Executors
  5. Methodology
  6. ToM-LM
  7. Model Fine-tuning
  8. Problem Prompting
  9. Problem Formalization
  10. External Deterministic ToM Reasoning
  11. Dataset
  12. Baselines
  13. DP
  14. SFGP
  15. DP$_{FT}$
  16. ...and 17 more sections

Figures (7)

  • Figure 1: Overview of ToM-LM, which consists of four principle stages: (1) Model Fine-tuning in which a given base model undergoes the fine-tuning process, (2) Problem Prompting in which a ToM problem and a one-shot example are constructed and given to the fine-tuned model, (3) Problem Formalization in which the model generates the corresponding symbolic formulation, and (4) External Deterministic ToM Reasoning in which the generated symbolic formulation is executed by SMCDEL model checker to give out the final result.
  • Figure 2: Output Distributions of DP, SFGP, DP$_{FT}$, and ToM-LM
  • Figure 3: Direct Prompting Template
  • Figure 4: Symbolic Formulation Generation Prompting Template
  • Figure 5: An example of premise, hypothesis, and symbolic formulation in MindGames.
  • ...and 2 more figures