Math-DB: A Discourse Framework for Mathematical Word Problems to Enhance LLM Reasoning

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

Math Discourse Bank (Math-DB) is a novel discourse framework and annotated dataset designed to enhance Large Language Model (LLM) reasoning for mathematical word problems. It addresses the issue of LLMs relying on statistical regularities rather than true logical reasoning, which causes performance drops with problem perturbations or irrelevant information. Inspired by the Penn Discourse TreeBank and mathematics education, Math-DB defines a hierarchy of discourse senses for quantitative reasoning, including categories like Change, Combine, Compare, and Equalize. The framework was applied to the GSM-Symbolic dataset of 12,500 problems, generating 47,815 sense-labeled discourse relations over 11,414 successfully-aligned instances, achieving a 91.3% pipeline yield. Experiments show that integrating Math-DB annotations into Chain-of-Thought (CoT) prompts consistently improves LLM performance across various difficulty levels.

Key takeaway

For NLP Engineers developing LLM solutions for mathematical reasoning, incorporating structured discourse frameworks like Math-DB into your prompting strategies is crucial. This approach moves models beyond statistical regularities, improving performance and robustness against problem perturbations or irrelevant information. Consider annotating problem datasets with quantitative discourse senses to enhance Chain-of-Thought prompting, leading to more reliable and logically sound mathematical problem-solving capabilities in your applications.

Key insights

Math-DB enhances LLM mathematical reasoning by providing a discourse framework that guides models beyond surface patterns to true logical understanding.

Principles

• LLMs often lack true logical reasoning.
• Discourse frameworks improve quantitative reasoning.
• Annotations can guide LLM problem-solving.

Method

Math-DB defines a discourse hierarchy with senses like Change, Combine, Compare, and Equalize. This framework is applied to mathematical problems to generate sense-labeled discourse relations, which are then incorporated into CoT prompts.

In practice

• Apply discourse senses to problem decomposition.
• Integrate structured annotations into CoT.
• Evaluate LLM robustness to problem perturbations.

Topics

• Large Language Models
• Mathematical Word Problems
• Chain-of-Thought Prompting
• Discourse Frameworks
• Dataset Annotation
• Quantitative Reasoning

Best for: AI Engineer, Machine Learning Engineer, Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Enhancing Science Classroom Discourse Analysis through Joint Multi-Task Learning for Reasoning-Component Classification — Automated discourse analysis using LLM augmentation improves utterance classification and reveals effective teaching patterns.
• PyraMathBench: Evaluating and Improving Mathematical Capability in Large Language Models — LLMs' mathematical failures stem from poor numerical computation and abstract numerical handling, necessitating integrated benchmarks and specialized modules.
• Construction of Knowledge Graph based on Language Model — Lightweight LLMs can construct Hyper-Relational Knowledge Graphs with performance comparable to GPT3.5.
• Knowing What to Solve Before How: Preplan Empowered LLM Mathematical Reasoning — Explicitly understanding "what to solve" before "how" significantly boosts LLM mathematical reasoning.
• MulDimIF: A Multi-Dimensional Constraint Framework for Evaluating and Improving Instruction Following in Large Language Models — A multi-dimensional framework and automated pipeline enhance LLM instruction-following evaluation and training.
• Recent LLMs can use filler tokens or problem repeats to improve (no-CoT) math performance — Recent LLMs can use filler tokens or problem repeats to boost math performance without explicit Chain-of-Thought.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.