THE QUESTIONS BEHIND THE SETTINGS
Experiments into how Maia chooses its moves, and what that means for playing on your phone.
Independent research by the Mobile Maia project. Read the findings here, then follow the evidence: methods, recorded games, datasets and reproduction tools are openly available.
START HERE · VISUAL GUIDE
What happens when you keep the opening variety and trim the least likely moves? Follow the worked Top-P example, opening comparisons and match results.
Read the visual guide
FOLLOW THE EVIDENCE
Full reports and reproducible evidence in the research repository.
The 79M comparison covers 3,000 full games, 5,400 opening rollouts and 180 quality positions at rating inputs 600, 1600 and 2400. Top-P 0.95 with full sampling through the first five moves improves the primary match score and preserves the measured broad opening mix.
Lower sampling settings made Maia stronger in the tested matches but reduced opening variety. The report covers the original 1.0/1.0, 0.5/0.9 and greedy comparisons, plus the 1.0/0.95 follow-up.
History improved average human-move prediction but strongly reinforced copying in a low-Elo Stonewall test. The recommendation is to keep history off by default; an experimental opt-in setting is proposed separately.
