EMNLP 2023

ATHENA: Mathematical Reasoning with Thought Expansion

JB. Kim, Hazel Kim, Joonghyuk Hahn, Yo-Sub Han

Presents ATHENA, an attention-based architecture that expands candidate mathematical thoughts step by step, improving math-word-problem solving under limited or varied training signals.

Overview

Math word problems reward careful intermediate reasoning, but models often collapse too early onto a single path. ATHENA is an attention-based architecture that expands candidate mathematical thoughts step by step.

By growing and comparing intermediate thoughts, ATHENA improves math-word-problem solving when training signals are limited or varied.

Why it matters

Better intermediate search can help models stay flexible before committing to an answer—especially in domains where a small early mistake cascades into a wrong final result.

Paper

Read the PDF on arXiv