Princeton's QUEEN: 4B Chess-Language Model Reaches Grandmaster Level

Adithya Bhaskar and colleagues from Danqi Chen's team at Princeton have released QUEEN, a chess language model: at only about 4 billion parameters, it reaches 2697 Elo on Lichess—close to the grandmaster (GM) median of 2730—and can explain its moves and position evaluations in natural language. The team describes it as a dual architectural and algorithmic innovation targeting the pain point that 'traditional engines play strong but stay silent, while language models can talk but play poorly.'

Confirmed

• QUEEN was released by the Princeton team; author Adithya Bhaskar considers it the deepest paper of his PhD
• Its strength reached 2697 Elo on Lichess, near the GM median of 2730; according to Elivnurotic's relay, there were still no signs of a plateau in strength when training stopped
• Core method: rather than having the model relearn chess from text, it plugs in an existing strong search engine as a 'silent expert,' repeatedly converting AlphaZero-style search processes into natural-language explanations and distilling them back into the model—a 'natural-language analogue of the AlphaZero algorithm'
• The model can generate natural-language explanations of its moves and plans, solving the problem of traditional engines 'giving moves but not reasons'

Why it matters

• Reproducing AlphaZero-level chess strength in a language model while retaining interpretability bridges the gap between 'strong engines' and 'explaining LMs'
• The training technique is, per the author, generalizable to other games, robot manipulation, computer use, and other sequential decision-making scenarios, with potential beyond chess
• Reaching grandmaster level at 4B parameters with no plateau yet suggests this distillation paradigm still has headroom at scale

2026-10-05 ~ 2026-10-07 · 7 related posts

• A natural-language AlphaZero: teaching LMs to play and explain chess — ChengleiSi · 2026-10-05
• [source] Queen: A 4B Chess LM Hits Grandmaster Level (2697 Elo) While Explaining Its Moves — Adithya Bhaskar · 2026-10-05
• QUEEN: a 4B chess-language model plays near GM strength and explains its moves in natural language — danqi_chen · 2026-10-05
• 4B chess language model plays at grandmaster level and explains its moves in natural language — WeijiaShi2 · 2026-10-06
• Princeton trains a 4B LLM to 2700 Elo in chess, with a technique transferable to robotics — Eliv_nurotic · 2026-10-06
• [source] Princeton's QUEEN: A 4B Model That Plays Chess at 2697 Elo and Explains Its Moves — 机器之心 · 2026-10-06
• [source] QUEEN paper distills AlphaZero-style chess search into language for LLMs — danqi_chen · 2026-10-07