Recent advances in reasoning models have improved performance in logic, mathematics, and programming but struggle with combinatorial reasoning tasks due to the vast search space. Existing neuro-symbolic techniques rely on formal translations and symbolic solvers but are limited and error-prone. We introduce CDCL-IC, a Chain-of-Thought (CoT) reasoning approach leveraging Conflict-Driven Clause Learning (CDCL) to prune search spaces via in-context learning. Applied to Sudoku, CDCL-IC significantly outperforms traditional CoT and o3-mini on 9x9 puzzles.
