Princeton's Chen Danqi Team Releases QUEEN: A 4B Model That Plays Chess at Master Level and Explains Moves

Chess engines long ago left humans far behind, yet they cannot say a single human sentence; frontier large models can write plausible analysis, but the moves they recommend may not hold up, making the explanation groundless. According to a report by Machine Heart, Princeton University's Danqi Chen team released the paper "Language Models that Play Chess and Explain Their Moves" (arXiv:2610.03695v1) on arXiv, and made the model and code public, attempting to connect these two ends.

The model they trained is called QUEEN, with about 4B total parameters, playing strength close to a typical grandmaster, while also being able to explain in natural language why it moves as it does. The paper's review found that most previously fine-tuned chess language models were trained and evaluated only on isolated tactical puzzles, and even in this restricted setting they struggled to select good moves, so the quality of their explanations was naturally limited.

The significance of this goes beyond the board. The report cites the paper as saying that as of October 2 there were 1,899 grandmasters worldwide, less than one in ten thousand active online players, and the vast majority of players will never get a grandmaster to sit down and explain chess to them in their lifetime. A model that plays well, explains clearly, and runs cheaply fills exactly this gap.

QUEEN's approach can be understood through a scenario: pair a silent master who can only point at the board with a commentator who is eloquent but mediocre at chess. The master is responsible for "seeing," the commentator for "speaking," and the key is to let the commentator truly understand the master's gestures.

The "silent master" is the Lc0 family's BT5 network (the paper calls it Leela), about 240M parameters and 15 layers, with a fixed input of 64 tokens, exactly corresponding to the 64 squares on the board, able to play at near-superhuman level without search. The "commentator" is HuggingFace's SmolLM3-3B, a general-purpose instruction model that has had almost no chess training.

The bridge between them borrows the gated cross-attention of the vision-language model Flamingo: the representation from Leela's layer i is fed in before layer 2i of the language model, with 16 bridge modules inserted in total, about 470M parameters. Unlike the original Flamingo, which takes only the encoder's last layer, this layer-by-layer connection lets the language model see both Leela's shallow and deep "thoughts." The gate initial value is set to zero, so the bridge is effectively absent at the start, then gradually opened during training to avoid early noise interference.

Next comes teaching the commentator to read the gestures. The team designed a four-stage question-and-answer curriculum, from shallow to deep: first static questions like "what piece is on a certain square," then dynamic questions like "where can this knight move" and "who can give check," then giving 1 to 8 moves and letting the model infer the future position before answering. All answers are generated and verified by program. In this stage both Leela and the language model itself are frozen, and only the bridge layers and the newly added vocabulary are trained.

The reason for adding dedicated tokens for 64 squares and 12 piece types is that the original tokenizer split square names inconsistently, and words like "bishop" and "knight" carry meanings unrelated to the board in everyday corpora. After training, the model's accuracy on the four types of questions reached 99.97%, 99.97%, 98.97%, and 96.07% respectively.