Language Models that Play Chess and Explain Their Moves
A 4B chess-language model called Queen is reported at 2697 Elo after iterative training.
The paper says Queen combines a silent chess encoder with an instruction-tuned language model, using cross-attention to pull chess concepts into explanations. Its training loop has the model analyze positions after candidate moves, consolidate that into a current-position explanation, then distill it back into the model. Across seven iterations, the authors report a rise from 1782 to 2697 Elo, beating frontier models on playing strength and puzzle accuracy in their tests. They also say its explanations are fluent and near GPT-5.6-Sol (high) on coherence by LM-based evaluation. HF Daily Papers' note
The paper says Queen combines a silent chess encoder with an instruction-tuned language model, using cross-attention to pull chess concepts into explanations. Its training loop has the model analyze positions after candidate moves, consolidate that into a current-position explanation, then distill it back into the model. Across seven iterations, the authors report a rise from 1782 to 2697 Elo, beating frontier models on playing strength and puzzle accuracy in their tests. They also say its explanations are fluent and near GPT-5.6-Sol (high) on coherence by LM-based evaluation. HF Daily Papers' note
score 4