Modern chess engines reach superhuman strength but offer no justification for their choices, while language models can generate plausible explanations yet play weakly. A research team introduces Queen, a 4-billion-parameter model designed to bridge this gap: playing chess at a typical Grandmaster level while explaining its moves and underlying plans in natural language.
The architecture uses an encoder-decoder design in which a specialized, silent chess encoder is connected to an instruction-tuned language model through cross-attention. This connection is trained using a question-answering curriculum aimed at extracting chess concepts embedded in the encoder's internal representations, allowing the model to verbalize knowledge that already exists in implicit form.
The core innovation lies in the iterative method used to refine explanations, described as a natural-language analog of the Bellman update from reinforcement learning. The model analyzes the positions that follow its top candidate moves, consolidates these analyses into an explanation of the current position, and then distills this synthesis back into its own training. Repeated over seven iterations, this loop raises the model's rating from 1782 to 2697 Elo, a gain of more than 900 points.
According to the authors, this level surpasses all current frontier models on both playing strength and puzzle-solving accuracy, despite Queen having roughly a thousand times fewer parameters than those general-purpose systems. Evaluations conducted using other language models further rate the fluency and coherence of Queen's explanations as close to those of an advanced reference model.
The authors suggest that this architecture, pairing a specialized silent expert with a language model through iterative distillation, could generalize to other domains where comparable expert encoders exist, such as games, robotics, or computer-use tasks. The significance of this work thus extends beyond chess: it offers a concrete method for making high-performing expert systems, which currently function as opaque black boxes, more interpretable.