Title: ChessMimic: Rating-Specific Transformer Models for Predicting Human Moves, Clock Usage, and Game Outcomes in Online Blitz Chess

Abstract:

This paper introduces ChessMimic, a framework comprising three distinct, lightweight encoder-only transformer architectures designed to predict human moves, thinking durations, and game results. These models are conditioned on board positions, recent move history, player ratings, and clock states. To optimize parameter efficiency while ensuring precise calibration across different skill levels, we train a separate instance of each model for every 100-Elo rating band.

Evaluations on a month-long, held-out dataset of Lichess Rated Blitz games demonstrate that ChessMimic’s human move prediction accuracy surpasses that of Maia-2 across all Elo categories. When compared to Maia-3, our 9-million-parameter model achieves accuracy levels comparable to the Maia-3-5M and Maia-3-23M variants, yet it does so without relying on the added complexity of Geometric Attention Bias.

Beyond move prediction, we developed a game outcome model that incorporates player ratings, time controls, and remaining clock times alongside the position. This model attained an out-of-sample AUC of 0.78, outperforming both Maia-2 and logistic regression baselines that rely on material balance, ratings, and clock time. Additionally, we trained a clock model to estimate human thinking times. Under ALLIE-style filtering criteria, this model yields a functional, though not state-of-the-art, per-ply thinking-time signal (Pearson r = 0.41, Spearman rho = 0.50, MAE 4.10 s), particularly when compared to ALLIE’s reported r = 0.70. The performance gap is primarily attributed to per-position bucket sharpness rather than issues with bucket-marginal calibration.

We provide a public demonstration at 1e4.ai and have open-sourced the code, per-band weights, and the C++ data-filtering pipeline on GitHub.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC