A Reasoning-Based Approach to Cryptic Crossword Clue Solving
This paper addresses the challenge of clue solving in British-style cryptic crosswords, where answers must simultaneously satisfy both a definition and a wordplay constraint. We propose the first LLM-based verifiable reasoning framework for this task. Our method employs a three-stage pipeline: (1) candidate answer generation, (2) derivation of a wordplay explanation, and (3) stepwise validation via a structured Python verifier. Crucially, we explicitly encode wordplay reasoning as executable, traceable Python operations—jointly modeling definition and wordplay pathways. Evaluated on the Cryptonite benchmark—a dataset of real cryptic clues—our approach achieves state-of-the-art accuracy, significantly outperforming prior methods. Every correct solution is accompanied by a complete, human-readable, and machine-verifiable reasoning chain. The implementation relies entirely on open-source LLM components, ensuring full transparency and reproducibility.