Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Table as Thought: Exploring Structured Thoughts in LLM Reasoning

Zhenjie Sun, Naihao Deng, Haofei Yu, Jiaxuan You

TL;DR

Table as Thought introduces a thought-level structured representation by organizing reasoning within a tabular schema, where rows encode sequential thoughts and columns capture constraints and context. The framework comprises a Schema Development Module, a Reasoning Verification Module, and a Table Construction Module to iteratively populate and verify reasoning tables, promoting self-consistency. Empirical results show substantial gains in constraint planning tasks and selective improvements in mathematical reasoning for capable models, though increased structural complexity can degrade performance for less capable models; a predefined schema often mitigates this degradation. The work bridges cognitive neuroscience insights with AI reasoning, highlighting the potential of structured internal representations to enhance planning and problem-solving in LLMs, while also outlining limitations in schema design for complex tasks and the need for broader model support.

Abstract

Large language models' reasoning abilities benefit from methods that organize their thought processes, such as chain-of-thought prompting, which employs a sequential structure to guide the reasoning process step-by-step. However, existing approaches focus primarily on organizing the sequence of thoughts, leaving structure in individual thought steps underexplored. To address this gap, we propose Table as Thought, a framework inspired by cognitive neuroscience theories on human thought. Table as Thought organizes reasoning within a tabular schema, where rows represent sequential thought steps and columns capture critical constraints and contextual information to enhance reasoning. The reasoning process iteratively populates the table until self-verification ensures completeness and correctness. Our experiments show that Table as Thought excels in planning tasks and demonstrates a strong potential for enhancing LLM performance in mathematical reasoning compared to unstructured thought baselines. This work provides a novel exploration of refining thought representation within LLMs, paving the way for advancements in reasoning and AI cognition.

Table as Thought: Exploring Structured Thoughts in LLM Reasoning

TL;DR

Table as Thought introduces a thought-level structured representation by organizing reasoning within a tabular schema, where rows encode sequential thoughts and columns capture constraints and context. The framework comprises a Schema Development Module, a Reasoning Verification Module, and a Table Construction Module to iteratively populate and verify reasoning tables, promoting self-consistency. Empirical results show substantial gains in constraint planning tasks and selective improvements in mathematical reasoning for capable models, though increased structural complexity can degrade performance for less capable models; a predefined schema often mitigates this degradation. The work bridges cognitive neuroscience insights with AI reasoning, highlighting the potential of structured internal representations to enhance planning and problem-solving in LLMs, while also outlining limitations in schema design for complex tasks and the need for broader model support.

Abstract

Large language models' reasoning abilities benefit from methods that organize their thought processes, such as chain-of-thought prompting, which employs a sequential structure to guide the reasoning process step-by-step. However, existing approaches focus primarily on organizing the sequence of thoughts, leaving structure in individual thought steps underexplored. To address this gap, we propose Table as Thought, a framework inspired by cognitive neuroscience theories on human thought. Table as Thought organizes reasoning within a tabular schema, where rows represent sequential thought steps and columns capture critical constraints and contextual information to enhance reasoning. The reasoning process iteratively populates the table until self-verification ensures completeness and correctness. Our experiments show that Table as Thought excels in planning tasks and demonstrates a strong potential for enhancing LLM performance in mathematical reasoning compared to unstructured thought baselines. This work provides a novel exploration of refining thought representation within LLMs, paving the way for advancements in reasoning and AI cognition.
Paper Structure (40 sections, 9 equations, 1 figure, 10 tables, 2 algorithms)

This paper contains 40 sections, 9 equations, 1 figure, 10 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Structures in LLM Reasoning
  4. Representations of Tables in LLM Inference
  5. Table as Thought
  6. Table as Thought.
  7. Schema Development Module
  8. Reasoning Verification Module
  9. Table Construction Module
  10. Experiments
  11. Tasks and Language Models
  12. Constraint Planning Tasks.
  13. Math reasoning tasks.
  14. Language Models.
  15. Text Thought Baselines
  16. ...and 25 more sections

Figures (1)

  • Figure 1: The Overall Pipeline for Table as Thought Reasoning. The figure illustrates how Table as Thought structures reasoning by iteratively populating a reasoning table based on the schema, verifying consistency, and updating the table until the final answer is achieved.