Queen: A 4B-Parameter Chess-Language Model That Plays at Grandmaster Level and Explains Its Moves

Forum topic · 小凯 · 2026-10-06

Summary

Researchers Adithya Bhaskar, Jeffrey Cheng, and Danqi Chen introduce Queen, a 4-billion-parameter chess-language model that combines grandmaster-level play with natural-language explanations of its moves and plans. The framework uses an encoder-decoder architecture that integrates a silent expert chess encoder with an instruction-tuned language model via cross-attention, trained through a question-answering curriculum to extract chess concepts from encoder representations. On top of this, an iterative distillation algorithm applies a natural-language version of Bellman updates: the model analyzes positions after its candidate moves, incorporates these analyses into explanations of the current position, and distills them back into the model. Over seven iterations, Queen's Elo improved by more than 900 points (from 1782 to 2697), substantially outperforming frontier models in both playing strength and puzzle accuracy despite roughly 1000x fewer parameters. Language-model-based evaluations show its explanations are fluent and approach the coherence of GPT-5.6-Sol. The approach offers a general recipe for applying language models to domains with silent expert encoders, such as games, robotics, and computer control.

Paper Overview

Field: NLP Authors: Adithya Bhaskar, Jeffrey Cheng, Danqi Chen Published: 2026-10-02 arXiv: 2610.03695

Abstract

Modern chess engines are silent experts: they play at a superhuman level, but do not offer explanations for their play. On the other hand, language models (LMs) can generate plausible-sounding explanations, but their weak playing strength limits the utility of their explanations.

The authors introduce Queen, a 4B-parameter chess-language model that can explain its moves and plans while playing at the level of a typical Grandmaster. The novel framework enables domain-specific reasoning through complementary components:

Encoder-decoder architecture: integrates a silent expert chess encoder with an instruction-tuned LM through cross-attention, trained via a question-answering curriculum to extract chess concepts from the encoder's representations.
Iterative distillation algorithm: uses a natural-language version of Bellman updates — the model analyzes the position after its optimal candidate moves and consolidates these analyses into an explanation of the current position, which is then distilled back into the model.

Results

After seven iterations, the model's Elo score improved by more than 900 points (from 1782 to 2697).
Queen substantially outperforms all frontier models in both playing strength and puzzle accuracy, despite having roughly three orders of magnitude fewer parameters.
Language-model-based evaluation shows the explanations are fluent and approach the coherence of GPT-5.6-Sol.

Significance

The generality of the architecture and training pipeline provides a path for applying language models to domains that have silent expert encoders, such as games, robotics, and computer operation.

---

*Auto-collected on 2026-10-06*