Faynt, a compact AI that selects controller inputs for Super Smash Bros. Melee, won 240 of 244 supported-character mirror games against 14 released AI agents. Here, a mirror means both sides used the same fighter from the rival’s roster; the score covers that matchup category, not the full test suite.
The team trained 10-million- and 75-million-parameter Transformer policies on about 840,000 ranked human replays. They emphasized stronger players’ examples, transferred behavior from the larger model to the smaller one, and used reinforcement learning—trial and error guided by game outcomes—on Fox mirror matches. The smaller model uses one set of trained weights across all 26 fighters.
The 244 supported-character games comprise 12 matches against each of five single-character specialists and two for each fighter in the rosters of nine multi-character agents. The wider evaluation had 1,312 games per Faynt model, with extended-roster and forced-mirror matches reported separately.
In the released comparisons, Faynt added no action delay while rival agents kept queues of 21 or 24 game frames. At 60 frames per second, those queues equal about 0.35 or 0.4 seconds. The authors did not isolate timing’s effect, so the score cannot show how much the result reflects the policies themselves versus Faynt’s faster action timing.
The authors have released the model weights, benchmark records and tournament software. The paper is still under review, and its results concern matches against these AI policies under the stated setup, not performance against human players.
Illustrative photo: Melee players pose at EVO 2014. Photo by vsmak350 via Wikimedia Commons, licensed under CC BY-SA 2.0 (resized for display). The photo does not show Faynt or its evaluation.