ChessMimic: Per-Rating Transformer Models for Human Move, Clock, and Outcome Prediction in Online Blitz Chess
arXiv:2606.04473v1 Announce Type: new Abstract: We present ChessMimic, a system of three small encoder-only transformers - for move, thinking-time, and outcome prediction - conditioned on the position, recent move history, player rating, and clock state. We fit a separate instance of each model per 100-Elo rating band, trading parameter efficiency for sharper per-skill calibration. On a held-out month-wide slice of Lichess Rated Blitz games ChessMimic's human move prediction accuracy outperforms Maia-2 in every Elo band. Compared to Maia-3, our 9M parameter model's accuracy sits between Maia-3
The continuous improvement in AI models for specific tasks, alongside advancements in transformer architectures, allows for specialized, high-performance applications like human behavior prediction in complex games.
This development showcases AI's increasing ability to model and predict nuanced human decision-making across varied skill levels, which has implications beyond chess for various predictive analytics and agentic systems.
A new benchmark for human move and outcome prediction in online chess has been established, leveraging per-rating band model specialization to achieve superior accuracy compared to prior state-of-the-art models.
- · AI researchers
- · Online gaming platforms
- · Behavioral analytics firms
- · AI model developers
- · Prior state-of-the-art AI models for game prediction
- · General-purpose AI models without specialized calibration
Improved AI performance in predicting human strategic behavior in competitive environments.
The methodology of 'per-rating' or per-skill calibration could be applied to other domains to enhance AI understanding and prediction of human performance.
This could lead to more sophisticated AI assistants or opponents that adapt precisely to individual human skill levels, enhancing training or competitive experiences across various sectors.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.LG