Language Models that Play Chess and Explain Their Moves

Imagine you're developing a chess AI that not only plays at a high level but also explains its moves in a way that makes sense to human players. Current chess engines can dominate the game but often leave users in the dark about their decision-making processes. On the flip side, language models can generate explanations, but their lack of chess expertise limits their usefulness. This creates a gap where players want to understand the reasoning behind moves but struggle to find a tool that can both play well and articulate its thought process clearly. This is what's called the interpretability problem in AI.

To address this, the authors propose a new model called Queen, which combines a chess-specific encoder with a language model through a unique architecture. The model learns to explain its moves by analyzing positions after making its top candidate moves, refining its explanations iteratively. This approach not only allows Queen to play at a Grandmaster level but also to provide fluent and coherent explanations that rival those of advanced language models. Compared to existing chess engines, Queen's ability to articulate its reasoning while maintaining high playing strength represents a meaningful step forward for AI in games, suggesting a broader application for similar frameworks in other domains like robotics and complex decision-making tasks.

The integration of chess expertise with language model capabilities is a significant advancement.

The model's performance is backed by substantial Elo rating improvements and coherent explanations.