應用研究
Otter Predicts Human Decisions Using Move History and Time Controls; 15.3M-Parameter Model Reports 55.23% Top-1 Accuracy
Rather than treating each board position as an independent sample, Otter incorporates the previous 20 moves and time-control information. The model outperformed Maia-2 on the authors’ Lichess rapid test, but was not directly compared with the published Maia-3 using the same data split.
Chess engines search for the objectively best move, while human behavior models aim to predict the move a player of a particular skill level will actually make. Otter’s key innovation is to reformulate the problem from single-board classification into sequence prediction conditioned on history and time. The 15.3M-parameter model processes the previous 20 moves, using a history encoder to capture opening preferences, position development, and behavioral inertia within the same game. A time-control module then adjusts the move distribution to reflect differences in decision-making under varying time pressure.
The authors constructed a dataset of 6.1 billion positions from 117 million Lichess rapid games and claim that training was completed in 30 days using a single T4 GPU. The model achieved a 55.23% top-1 move-matching accuracy and a 90.95% top-5 accuracy. Across 11 Elo brackets, the 1900–1999 group recorded the highest top-1 accuracy, at 57.38%. These results suggest that, for modeling human behavior, prior actions and time constraints may offer more value than simply scaling up a position-only model. The approach could also be useful for skill-adjusted training partners, mistake prediction, and analysis tools centered on the question, “What would a similarly rated player do?”
However, claims that Otter “surpasses the previous state of the art” should be interpreted cautiously. The paper primarily compares Otter with Maia-2. Maia-3, released earlier this year, reported a 57.1% move-matching accuracy and offers 5M-, 23M-, and 79M-parameter models. However, the two systems use different game formats, time controls, data periods, and test splits, so their 55.23% and 57.1% results cannot be directly compared. Otter’s dataset size also does not guarantee reproducibility: researchers still need to examine deduplication, player leakage, missing time fields, and distribution drift across different months. The most valuable next step would be to evaluate Otter, Maia-3, and other human move models on the same public test set, with separate ablation studies for history length and time-control signals.