Gaming Retrieval Augmented Language Models for Agentic, Thoughtful Wordplay

Abstract

Agentic-like uses of language models for domain-specific tasks have grown substantially on the precipice of rapidly improving closed and open-source pre-trained models. Techniques such as retrieval augmentation further enhance model performance by efficiently imputing relevant context from which to perform prompt completion. However, there are still elusive goals in this research area: producing a convenient model architecture that is adaptable to ever-shifting generation of foundational model; AI systems that are highly interpretable by human observer; and performing well in games that require nontrival thinking executive decision making to solve the problem. I approach all three problems through the development of an agentic cryptic crossword solver. A decision model is implemented that carefully resembles that of a human crossword solver: retrieving relevant context, generating thoughts, voting on the best thoughts, and making executive decisions though the game. A standalone application is also developed the visualise this process. To put this solver into practice: I leverage state-of-the art closed language models, local ensemble models, dense retrieval, two-stage fallback online retrieval, reranking, and LM-supervised contextual learning. Without relying on validity of the crossword board or domain specific fine tuning, the model effectively solves crosswords through seemingly coherent justification and decision making.