QUEEN: 4B Chess Model Matches Grandmasters and Explains Moves in Natural Language

Introduction

Chess engines have long surpassed human players but remain silent, outputting only a move and a win-rate bar. Language models can articulate reasoning but often play poorly. A team led by Danqi Chen at Princeton University bridges this gap with QUEEN , a 4-billion-parameter model that reaches near-grandmaster playing strength (2697 Lichess Elo) and explains its moves in coherent natural language.

Background: The Silence of Engines, the Weakness of Talkers

From Shannon's 1950 vision to Deep Blue, AlphaZero, and Leela Chess Zero (Lc0), chess engines have grown stronger yet still produce only a move and a numeric evaluation. Fine-tuned chess language models have mostly been evaluated on isolated tactical puzzles and struggle to select good moves. Frontier models like GPT-5.6-Sol can generate plausible explanations, but running a full game at high reasoning effort costs over $15. With only 1,899 grandmasters worldwide — far less than one per ten thousand active online players — a model that plays well, explains clearly, and runs cheaply fills a significant void.

Architecture: A Silent Master and a Talkative Commentator

The system pairs two frozen pre-trained models:

Silent master: Leela BT5 network (Lc0 family), ~240M parameters, 15 layers. Input is 64 tokens corresponding to the 64 squares; it reaches superhuman strength without search.

Talkative commentator: Hugging Face's SmolLM3-3B, a general instruction model with almost no chess-specific training.

A bridge inspired by Flamingo's gated cross-attention connects them. Leela's layer i representations are inserted before the language model's layer 2i , adding 16 bridge modules (~470M parameters). Unlike Flamingo, which only uses the encoder's final layer, this layer-wise connection lets the language model see both shallow and deep Leela representations. Gating is initialized to zero (bridge effectively absent) and opens gradually during training to avoid early noise.

Stage 1: Teaching the Commentator to Read the Master's Gestures

A four-stage question-answer curriculum, from static to dynamic reasoning, trains only the bridge layers and a new chess vocabulary (64 square tokens + 12 piece tokens) because the original tokenizer splits square names inconsistently and words like "bishop"/"knight" carry non-chess meanings. Leela and the language model stay frozen.

Static: "What piece is on square e4?"

Dynamic: "Where can this knight move?" / "Who can deliver check?"

Sequential: Given 1–8 moves, predict the resulting position and answer questions about it.

All answers are programmatically generated and verified. After training, accuracy on the four question types reaches 99.97%, 99.97%, 98.97%, and 96.07% respectively.

Stage 2: Self-Improvement via Natural-Language Bellman Update (Iterative Search Distillation)

Inspired by AlphaZero's search-and-distill loop, QUEEN performs the same idea in natural language:

For a position, the model proposes three candidate moves.

It plays each move and re-analyzes the three resulting positions.

If a sub-analysis contains an error, it drills down recursively until reaching the model's limit.

Qwen3.8-27B merges the three sub-analyses into a single explanation for the original position, explicitly stating why the chosen move is better and forbidding new content.

Only if the merged conclusion improves upon the model's original judgment is the data kept for the next training iteration.

Example: White candidates Rh2, Qh2+, Qxf3. From the opponent's perspective, the three resulting positions evaluate to forced mate, -0.1, and 0.0. Taking the move worst for the opponent (Rh2) reveals the quiet move is winning, while the tempting check Qh2+ lets the king escape. Each iteration starts from ~400,000 root positions, yielding 150,000–280,000 training samples. After seven iterations, the eighth generation is named QUEEN.

Experimental Results

Playing Strength

Each model plays 32 games against eight engines of varying strength, converted to Lichess Elo:

QUEEN: 2697

Gemini-3.1-Pro: 2201

GPT-5.6-Sol: 2071

Previous 4B model C1-4B: 0–32 loss

Reference: Lichess grandmaster blitz median is 2730. Over seven iterations, QUEEN's Elo rose from 1782 to 2697 (+915), with a temporary ~100-point drop between iterations 5 and 6.

Reasoning Ability

Tested on 1,000 tactical puzzles and 1,000 practical positions:

Tactical puzzles: QUEEN's first-move non-blunder rate 91.6% , 9 percentage points higher than Gemini.

Practical positions: non-blunder rate 97.1% .

Weaknesses

Judged by GPT-5.6-Sol, QUEEN scores 3.51 on structural coherence (on par with Sol) but only 2.76 on conceptual coherence (vs. Sol's 3.61), with slightly lower fluency. It finds correct variations but misapplies terms like "pin" and "outpost". The authors analyze that self-distillation compresses the search tree into weights but cannot teach description of unseen patterns; errors may propagate during merging.

Ablation Studies

Removing either the Leela encoder or the four-stage curriculum drops Elo to 514.

QUEEN is not merely copying Leela: in positions with ≥5 reasonable moves, it chooses differently 53.8% of the time.

Alternative Initialization

Using Stockfish handcrafted features + templates for initial explanations (no frontier LLM) yielded higher early Elo (2115 at iteration 1).

Conclusion: A Transferable Recipe

QUEEN demonstrates a general recipe: in any domain with a "silent expert" model, connect it to a language model via cross-attention, teach the language model to read the expert's representations through a curriculum, then iteratively improve explanations via search-and-distill. The authors suggest applicability to games, robotics, and computer control — areas rich in strong but non-verbal specialist models. As the next world championship broadcast in Geneva still shows only a numeric evaluation bar, an AI that turns those numbers into reasoning may not be far off.