Publication Date

Spring 2026

Degree Type

Master's Project

Degree Name

Master of Science in Computer Science (MSCS)

Department

Computer Science

First Advisor

Robert Chun

Second Advisor

Thomas Austin

Third Advisor

Avinash Sharma

Keywords

Transformers, Stepwise Deduction, Constraint Learning, Sudoku

Abstract

Sudoku is a constraint satisfaction problem that serves as a testbed for studying reasoning and stepwise deduction. While many solvers can produce correct solutions, they often fail to generate human interpretable sequences of logically consistent steps. This study investigates whether a transformer trained on stepwise deduction traces can learn to solve Sudoku puzzles through sequential, logically deducible moves. Results show that solve accuracy improves significantly with more training data, reaching approximately 78% at 1500k samples. Stepwise analysis indicates that the model effectively learns simple strategies, achieving near-perfect performance. However, performance on more complex strategies remains limited. Overall, while the model demonstrates strong constraint awareness and produces largely deducible steps, it struggles to consistently capture more complex reasoning, highlighting the need for improved data balance or architectural enhancements.

Download

Share

COinS