Training a Chess Reasoning Agent (blog.lukesalamone.com)

A developer has successfully trained a language model reasoning agent to play chess, utilizing models with sizes ranging from 0.8 billion to 9 billion parameters. This innovative approach leverages specialized tools to keep track of the chessboard state, drastically improving performance and reducing issues such as confabulation that smaller models typically face. The researcher emphasizes explainability and intuitive gameplay over traditional performance metrics, marking a shift in how chess engines could operate. The significance of this development lies in its application of a "fractally verifiable" framework, where every move and justification can be distinctly verified, enhancing understanding and engagement for players. Notably, the agent uses atomic tools that align with human reasoning, like tracking piece positions and available legal moves, while requiring the model to provide justifications for its actions. With demo code available on GitHub, this project not only advances the AI/ML community's exploration of reasoning capabilities in games but also sets a precedent for creating explainable AI systems in complex environments.