Faynt: Scaling and Optimizing Policies for Competitive Melee
Faynt's 10M Transformer wins 98.4% of same-character Super Smash Bros. Melee games against fourteen prior releases.
Faynt is a family of 10M- and 75M-parameter Transformer policies for Super Smash Bros. Melee, each controlling all 26 characters from one checkpoint. After reinforcement learning, the 10M model wins 240 of 244 same-character games (98.4%) against fourteen prior releases, though those opponents keep 21- or 24-frame action delays that Faynt does not use. It also wins all 68 games against a privately supplied zero-delay Slippi-AI model. Pretraining uses about 840,000 human replays; post-training adds curricula, 75M-to-10M distillation, and Fox-only RL. Inference averages 5.2 ms (10M) and 8.7 ms (75M) per decision on an NVIDIA T4, and weights plus benchmarks are open-sourced.
- One 10M or 75M checkpoint controls all 26 Melee characters.
- The 10M policy wins 240 of 244 same-character games, or 98.4 percent.
- Opponents retain 21- or 24-frame delays; Faynt adds none, and that gap is unisolated.
- Pretraining uses about 840,000 human replays, then distillation and restricted RL.
- Weights, both benchmarks, and an automated tournament platform are open-sourced.
- Vendors
- NVIDIA
- Products
- Super Smash Bros. MeleeNVIDIA T4
From the original article
1 min read at arxiv.orgWe introduce Faynt, a family of 10M- and 75M-parameter Transformer policies for Super Smash Bros. Melee, each controlling all 26 characters with a single checkpoint. After reinforcement learning (RL), the 10M wins 240 of 244 same-character games (98.
By Ali Janati · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.