Faynt: Scaling and Optimizing Policies for Competitive Melee
This paper tackles competitive Super Smash Bros. Melee with a single Transformer policy that controls all 26 fighters from one checkpoint. The authors pretrain 10M- and 75M-parameter models on roughly 840k human replays, then apply rank-based supervised curricula, 75M-to-10M distillation, and Fox-only reinforcement learning. The resulting 10M policy wins 240 of 244 benchmark games against prior releases and runs within the 16.67 ms frame budget, while the authors openly acknowledge key confounds in evaluation.
The paper makes a solid engineering contribution backed by an unusually systematic training pipeline and honest reporting of limitations. The 10M RL policy achieves a 98.4% win rate on supported characters and wins all 68 zero-delay games against a private Slippi-AI checkpoint, which is impressive. However, the headline comparisons against publicly available opponents are heavily confounded because those opponents retain 21-24 frame action delays while Faynt uses none, and the authors admit they have not isolated the effect of this difference. Consequently, it is hard to tell how much of the margin comes from faster reactions versus better strategic play.
The pretraining scaling study is methodical: the authors fit compute-optimal scaling laws for controller-prediction loss, run successive halving to transfer optimizer settings across sizes, and release all weights and benchmark code. The inference stack is also carefully engineered; by removing host synchronizations and replaying CUDA graphs, the 10M policy averages 5.2 ms per decision on a T4 with a 99th-percentile latency of 6.8 ms. A useful empirical finding is that weighted validation loss on high-ranked winners predicts win-rate ordering better than overall negative log-likelihood.
The biggest limitation is the reaction-delay confound. Because most public opponents queue actions for 18-24 frames while Faynt acts every frame with zero added delay, the reported margins cannot be read as pure skill gaps.
A second issue is that RL training is restricted to Fox mirror matches, meaning the other 25 characters share the updated weights but receive no RL experience. While both models still win on other characters, the 75M loses badly to several specialists on their native fighters, suggesting that Fox-only RL does not robustly transfer to the full roster for the larger model. Additionally, the two sizes follow different RL histories—the 10M requires a second leash run with altered hyperparameters—so size comparisons are entangled with training recipe.
The evidence supports the claim that supervised post-training and distillation improve play, but not always through lower overall NLL. The supervised 10M wins 69.7% versus 45.4% for the pretrained 75M despite higher overall held-out controller-prediction loss, which is a nice demonstration that imitation loss and playing strength can diverge. However, comparisons to prior work are uneven: MIMIC and Slippi-AI are evaluated with different action delays and conditioning schemes, and some opponents used during RL checkpoint selection reappear in the final benchmark, so those results are not independent. The zero-delay evaluation against the private Slippi-AI checkpoint is cleaner, but with 68 games total, the sample is modest.
Reproducibility is strong on artifacts—weights, benchmark suites, and tournament software are open-sourced—but reproducing the training pipeline from scratch is demanding. The full corpus has 839,942 replays, RL requires 96 Dolphin emulators on 32 CPU cores, and the final runs used hundreds of dollars of GPU time with extensive successive-halving searches. Hyperparameters are documented in depth, yet the final RL checkpoint for the 10M was chosen after looking at win rates against opponents that later appear in the benchmark, and the private zero-delay Slippi-AI checkpoint is not publicly available. Independent researchers will be able to evaluate the released checkpoints but may struggle to replicate the full training trajectory.
We 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 (98.4%) against fourteen specialist and multi-character releases on their supported rosters, with a winning record against every release. These opponents retain 21- or 24-frame action delays; Faynt uses no added delay, and we have not isolated the effect of this difference. In a separate evaluation against a privately supplied zero-delay Slippi-AI model, the 10M wins all 68 games across two conditioning settings. We study architecture, optimization, scaling, and hyperparameter transfer to guide pretraining on approximately 840,000 human replays. Post-training combines rank- and outcome-based curricula, 75M-to-10M distillation, and RL restricted to Fox mirror matches. On the initial 152-game benchmark, the supervised 10M wins 69.7% of games, compared with 45.4% for the pretrained 75M, despite higher overall held-out controller-prediction loss. The weighted validation loss used for supervised checkpoint selection agrees with the win-rate ordering of all four pretrained and supervised policies. After supervised post-training, both models take less damage per minute, build larger early leads, and win more often after losing the first life. Optimized inference on recorded game states averages 5.2 ms per decision for the 10M and 8.7 ms for the 75M on an NVIDIA T4, excluding emulator execution and communication. We open-source the weights, both benchmark suites, and a platform for automated model tournaments.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.