Language Models that Play Chess and Explain Their Moves

Language Models that Play Chess and Explain Their Moves. By Adithya Bhaskar and two others, from Princeton University. Published on the second of October. Its code repository is worked out: about one hundred and thirty source files, with a readme and a licence.

Modern chess engines are silent experts. They play at a superhuman level and offer no explanation for any of it. Language models sit at the other end: they can produce explanations that sound plausible, but they play weakly, and an explanation from a weak player is of limited use. One side has the strength, the other has the words, and neither has both. What the authors want is a single model in which the two come together: one that plays well enough for its reasoning to matter, and speaks well enough for that reasoning to be heard.

Queen is a chess-language model of four billion parameters that explains its moves and plans while playing at the level of a typical Grandmaster. The framework behind it has two complementary parts: an encoder-decoder architecture that gives the language model access to chess expertise, and an iterative distillation algorithm that improves what it says.

The architecture connects a silent expert chess encoder to an instruction-tuned language model through cross-attention. It is trained with a question-answering curriculum to extract chess concepts from the encoder's representations, so the language model learns to read what the encoder knows. The encoder stays the expert. The language model becomes its interpreter, and the curriculum of questions and answers is how the two learn a shared vocabulary of chess concepts. This is the domain-adapted model that the next step starts from.

The explanations are then improved by a natural-language analogue of the Bellman update. The model analyses the positions that follow its top candidate moves, consolidates those analyses into an explanation of the current position, and that explanation is distilled back into the model. Each round, the explanation of a position is built from looking one move ahead. The analogy to the Bellman update is that the value of where you are is assembled from the value of where you could go next. Here the quantity being passed backwards is not a number but an explanation in plain language.

Over seven iterations, the model gains more than nine hundred Elo points, rising from one thousand seven hundred and eighty-two to two thousand six hundred and ninety-seven. It substantially surpasses all frontier models on both playing strength and puzzle accuracy, despite having three orders of magnitude fewer parameters. Evaluations by language models find its explanations fluent, and approaching GPT five point six Sol, at its high setting, in coherence.

The authors present this as a recipe rather than a chess result alone. Wherever a silent expert encoder already exists, and they name games, robotics and computer use, the same pairing of encoder and language model, followed by look-ahead distillation, could give that expert a voice.