Chess engines have long outplayed humans, but engines like Stockfish and LC0 are silent experts: they tell you the best move, not why it is best. Language models are the opposite — fluent explanations backed by weak play. Queen, a 4B-parameter chess-language model from the Princeton NLP team, tries to glue the two halves together: it plays at the level of a typical Grandmaster while explaining its moves and plans.
Giving the Silent Expert a Voice
Queen uses an encoder-decoder architecture. The core idea is to take chess's silent expert — the LC0 engine — as an encoder and plug it into an instruction-tuned language model through cross-attention. Training follows a question-answering curriculum that teaches the LM to read chess concepts out of the encoder's representations. Think of it as letting a talkative model sit in on a master's inner monologue, then retell it in its own words.
A Natural-Language Bellman Update
What really sends the playing strength up is iterative self-distillation. The team built a natural-language analog of the Bellman update from reinforcement learning: the model analyzes the positions after its top candidate moves, consolidates those analyses into an explanation of the current position, and distills that explanation back into itself. Over seven iterations, its Elo climbed from 1782 to 2697 — a gain of more than 900 points. Per the authors' own evaluation, it surpasses all frontier models on both playing strength and puzzle accuracy with three orders of magnitude fewer parameters, and LM-based evaluations put its explanation coherence close to GPT-5.6-Sol (high). Note these comparison numbers all come from the authors' own benchmarks; no third-party replication exists yet.
Open-Sourced, Still Rough
The code is on GitHub under Apache 2.0, with models on Hugging Face under the princeton-nlp collection. But the README itself admits the "extended README and clean code release" is still coming: the repo has just 5 commits, and the vLLM setup required for GPU inference is "still being validated." Anyone hoping to replicate immediately should calibrate expectations.
So What
The thing worth remembering from this paper is not chess — it is a recipe. Anywhere a silent expert encoder exists — games, robotics, computer use — you can bolt domain-specific reasoning onto a language model via expert representations, cross-attention, and iterative self-distillation. If a 4B model beating generalist models three orders of magnitude larger keeps working in vertical domains, many use cases no longer need to wait for the next, bigger general model.
Reference: arXiv:2610.03695 (https://arxiv.org/abs/2610.03695); code at github.com/queen-project/queen