Faynt: Scaling and Optimizing Policies for Competitive Melee
A single 10M-parameter Transformer policy beat every tested Melee bot release in same-character matches.
Faynt controls all 26 characters from one checkpoint, rather than using character-specific models. In the paper’s main same-character benchmark, the 10M model won 240 of 244 games, though the authors note the opponents had built-in action delays and Faynt did not. Against a privately supplied zero-delay Slippi-AI model, it won all 68 games in a separate test. The authors also report open-sourcing the weights, benchmarks, and automated tournament platform. ArXiv · AI/CL/LG's note
Faynt controls all 26 characters from one checkpoint, rather than using character-specific models. In the paper’s main same-character benchmark, the 10M model won 240 of 244 games, though the authors note the opponents had built-in action delays and Faynt did not. Against a privately supplied zero-delay Slippi-AI model, it won all 68 games in a separate test. The authors also report open-sourcing the weights, benchmarks, and automated tournament platform. ArXiv · AI/CL/LG's note
score 5