Independent research supporting design decisions in Mobile Maia and Mobile Maia Preview.
Research reports, datasets and reproduction tools live in this repository, separate from application source and release distributions. App READMEs contain brief decision summaries and links here.
Updated research into best playing settings for Mobile Maia
A visual argument for opening variety on, Temperature 1.0 and Top-P 0.95: keep the opening mix, reduce the chance of costly mistakes, and improve later play. Includes a worked Top-P example, opening comparisons, move-quality results and match scores, with links to the full study and recorded data.
Read the report · Evidence and reproduction guide
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. It is recommended as a balanced default candidate in a separate app change; finer opening variation and human-move fidelity remain unresolved. Application defaults are unchanged.
Read the report · Evidence and reproduction guide
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. Recorded games, human opening references, diagnostics, figures and reproduction tools are included. Results describe these experiments, rather than a universal Elo guarantee or an upstream recommendation.
The original app-repository publication remains available for older app versions.
Read the report · Evidence and reproduction guide
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.
The study includes 20,160 main human-game positions, 4,000 opening positions, 384 tactical diagnostics and 1,000 targeted Stonewall opening trials. It does not establish a change in full-game Elo or claim an explicit upstream recommendation against history.
This is independent Mobile Maia research, not an official Maia-team publication. Sources, model hashes and methodological limits are documented per study. Original code and documentation are licensed under AGPL-3.0-only; third-party models, datasets and software retain their upstream terms.