Elo-Disentangled Player-Style Embeddings for Human Chess via Rating-Conditioned Residual Move Model

Researchers developed a machine learning approach that separates chess playing strength (Elo rating) from individual player style by using a rating-conditioned base model combined with learned player embeddings. The method achieves 27-37% relative improvement in move prediction accuracy over existing models while successfully disentangling stylistic preferences from playing skill level.

This research addresses a fundamental problem in representation learning: isolating stylistic patterns from confounding variables like skill level. By decomposing player behavior into a rating-typical component (what an average player of that strength would play) and an Elo-orthogonal stylistic component, the researchers create interpretable embeddings that capture individual chess style independently of strength.

The technical innovation lies in the residual formulation using Maia chess models and Stockfish engine features as anchors. The base model shows that engine analysis becomes increasingly valuable at higher skill levels—negligible below 1200 Elo but contributing 0.085 nats at 2800+ Elo—revealing how human play differs systematically across the rating spectrum. The 68% top-1 move-matching rate on the benchmark represents a 33% relative improvement over prior work, with gains concentrated at elite levels where style variation peaks.

While the player embeddings add minimal direct move-prediction value, their representational power proves significant for generalization and player re-identification from unseen games. The critical finding—that linear rating probes achieve only R² = 0.06 from the embeddings—provides empirical evidence of successful disentanglement, distinguishing this from approaches that merely compress rating information alongside style.

This work has implications beyond chess, suggesting practical methods for learning human-specific representations in domains where skill and style are correlated. The economical alternative to fine-tuning per-player models could inform personalization systems in games, sports analytics, and human-AI interaction design where understanding individual patterns matters alongside aggregate performance metrics.

• →A residual model using rating-conditioned baselines successfully disentangles player style from Elo rating, achieving R² = 0.06 when predicting rating from learned embeddings.
• →The augmented base model improves move prediction by 27-37% over Maia-3 alone, with Stockfish features showing monotonic value gains across the rating spectrum.
• →Player embeddings generalize to held-out decisions and enable above-chance player re-identification from disjoint games without overfitting.
• →Stockfish analysis is negligible for sub-1200 players but contributes substantially (+0.085 nats) at elite levels above 2800 Elo.
• →The approach offers an interpretable, computationally efficient alternative to per-player fine-tuning for capturing individual behavioral patterns in skill-stratified domains.