Language Models That Play Chess (2700 Elo) and Explain Their Moves (arxiv.org)

A groundbreaking development in AI has been announced with the introduction of "Queen," a 4-billion-parameter chess-language model capable of playing at a Grandmaster level (around 2700 Elo) while providing explanations for its moves. Unlike traditional chess engines, which operate silently, Queen leverages a unique framework combining an encoder-decoder architecture with an iterative distillation algorithm. This allows it to enhance its performance and produce fluent, coherent explanations by integrating a chess encoder with an instruction-tuned language model trained through a question-answering curriculum. Over multiple iterations, the model improves its strength by over 900 Elo points, significantly outperforming other models despite having significantly fewer parameters. This innovation is crucial for the AI/ML community as it demonstrates the potential for language models to operate effectively in specialized domains, like chess, where silent expert systems exist. The implications extend beyond chess; the architecture and training methods used in Queen can be adapted to other fields such as robotics and complex problem-solving, allowing AI systems to not only perform tasks but also articulate their rationale. This paves the way for more transparent and interpretable AI applications, enhancing user trust and facilitating better human-AI collaboration.