AI model separates chess style from skill with 33% better move prediction
New embedding learns what makes your chess style unique, independent of your rating.
Chess AI just got a personality upgrade. Researcher Jason Carlson from arXiv presents a novel method that learns a player's individual style completely separate from their playing strength (Elo). The approach uses a 'rating-conditioned residual' framework: a base model predicts what a typical player of a given rating would play, using Maia-3 policy logits and Stockfish-derived features scored over candidate moves. This base alone already crushes existing models — achieving 27–37% relative improvement in negative log-likelihood across all Elo levels, and lifting top-1 move accuracy from 0.51 (Maia-2) to 0.68, a 33% relative gain.
On top of this strong base, a compact per-player embedding vector (z) captures only the deviations from typical play. This vector adds almost nothing to raw move prediction but shines representatively: it can re-identify players from disjoint games above chance and, crucially, a linear probe predicts Elo from z with only R²=0.06 — strong evidence that style truly lives on an orthogonal axis to skill. The paper argues this is an economical, interpretable alternative to per-player fine-tuning of large models, offering a way to analyze chess styles at scale.
- Base model (Maia-3 + Stockfish) improves move prediction by 27-37% NLL across all Elo ratings, with top-1 accuracy rising from 0.51 to 0.68.
- Per-player embedding captures stylistic deviations from rating-typical play, re-identifying players from disjoint games above chance.
- Linear probe on embedding predicts Elo with only R²=0.06, confirming true disentanglement of style from playing strength.
Why It Matters
Enables scalable analysis of chess style independent of skill, with implications for coaching, player identification, and AI personalization.