Competitive real-time fighting game AI has long been a rigorous testbed for reinforcement learning and policy generalisation. Super Smash Bros. Melee, with its 52-character roster, frame-perfect inputs, and continuous state space, presents a particularly demanding benchmark. Makhas and Sabanis' Faynt introduces a family of Transformer policies—10M and 75M parameters—each controlling all 26 characters with a single checkpoint. The results are striking: after reinforcement learning, the 10M parameter model wins 240 of 244 same-character games (98.4%) against fourteen specialist and multi-character releases, achieving a winning record against every opponent. Against a privately supplied zero-delay Slippi-AI model, the 10M wins all 68 games across two conditioning settings. These results establish Faynt as a strong generalist Melee player while raising important questions about the role of action delays in AI evaluation.
From Pretraining to Generalisation
The Faynt pipeline begins with supervised pretraining on approximately 840,000 human replays. This diverse dataset provides coverage across character matchups, skill levels, and game states. The architecture is a Transformer policy that takes the current game state as input and outputs controller actions. After pretraining, the authors apply a combination of rank- and outcome-based curricula, 75M-to-10M model distillation, and reinforcement learning restricted to Fox mirror matches. This post-training regimen is critical: the authors show that on the initial 152-game benchmark, the supervised 10M policy wins 69.7% of games, compared with only 45.4% for the pretrained 75M model—despite the 75M model having lower 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, suggesting that prediction loss alone is not a reliable proxy for gameplay strength.
What makes this particularly interesting is the scaling dynamics. The 75M model starts with an advantage in raw predictive performance, but the 10M model surpasses it after targeted post-training. The distillation step (75M-to-10M) appears to transfer the stronger pretrained capabilities into a more efficient model, while the Fox-only RL fine-tuning further specialises the policy for competitive play. The fact that the 10M model, even before RL, outperforms the 75M model after supervised training alone demonstrates that larger parameter count does not automatically translate to better gameplay when the training signal is indirect (controller prediction rather than win/loss outcomes).
Reinforcement Learning and Curriculum Design
The reinforcement learning phase is deliberately narrow: Fox mirror matches only. This focused approach avoids the credit-assignment problem that arises when RL is applied across all 26 characters simultaneously, and it allows the policy to develop character-specific techniques that generalise through the shared Transformer architecture. The post-training combines rank-based curricula (progressing from easier to harder opponents) and outcome-based curricula (focusing on states where the policy loses the first life). After this post-training, both the 10M and 75M models exhibit measurable behavioural improvements: they take less damage per minute, build larger early leads, and win more often after losing the first life. These are precisely the metrics that matter in Melee, where position and lead preservation often determine game outcome.
The curriculum design reflects a careful understanding of Melee's learning curve. Early opponents are selected to reinforce fundamental techniques (spacing, combo execution, shield safety), while later opponents introduce more advanced strategies (edge-guarding, disadvantage state recovery). Outcome-based curriculum further ensures the policy learns to recover from adverse positions—a common scenario in Melee where the first life is often lost in neutral. The success of this curriculum is evident in the post-training results, where the models show improved clutch performance.
Benchmark Results and Comparative Evaluation
The evaluation protocol is comprehensive. The initial 152-game benchmark pits policies against a diverse set of opponents, including specialist Melee AIs and multi-character releases. The 10M model's 69.7% win rate on supervised policy and 98.4% win rate after RL against same-character opponents are particularly notable. The authors also evaluate against a zero-delay Slippi-AI model, where the 10M wins all 68 games. This result is significant because it demonstrates the 10M policy's strength even when the opponent has no added action delay—a factor the authors acknowledge but do not isolate as a causal variable.
The paper thoughtfully addresses the action-delay confound. The specialist and multi-character releases retain 21- or 24-frame action delays, while Faynt uses no added delay. The authors explicitly state they have not isolated the effect of this difference, which is a honest and important caveat. The zero-delay evaluation against the Slippi-AI model partially addresses this, but the causal contribution of delay versus policy capability remains an open question. Future work could systematically vary the action delay to quantify its impact.
Inference Efficiency and Practical Deployment
One of the paper's practical contributions is the inference timing benchmark. On an NVIDIA T4, the 10M policy averages 5.2 ms per decision and the 75M policy averages 8.7 ms, excluding emulator execution and communication. These are remarkably low latencies for Transformer policies operating on full Melee state spaces. The 10M model's latency is particularly impressive given its broader character coverage (26 characters vs. character-specific policies). For context, human reaction times are typically in the 100–250 ms range, so these policies operate at speeds comparable to human perception-action loops, well within the 60 fps (16.7 ms per frame) Melee frame rate. The efficiency gains from model compression (75M → 10M) enable real-time deployment without specialized hardware, making Faynt accessible for tournament play, replay analysis, and bot development.
Open-Source Infrastructure
The authors open-source the model weights, both benchmark suites, and a platform for automated model tournaments. This infrastructure is perhaps as valuable as the algorithmic contributions. The benchmark suites provide standardised evaluation protocols that the community can use to compare new Melee AIs. The tournament platform enables head-to-head competitions with automated result recording, win-rate calculation, and bracket generation. By making these resources publicly available, the authors lower the barrier to entry for Melee AI research and enable reproducible comparisons across institutions and hardware configurations.
Scaling, Transfer, and Future Directions
The paper provides a detailed study of architecture, optimization, scaling, and hyperparameter transfer. The pretraining on 840,000 human replays serves as a strong foundation, and the subsequent fine-tuning demonstrates that targeted post-training can extract substantial performance gains from moderate-scale models. The scaling investigation—comparing 10M and 75M models—suggests that for this task, model efficiency and targeted fine-tuning matter more than raw parameter count. The hyperparameter transfer from controller-prediction to reinforcement learning is another area worth exploring: the authors note that held-out prediction loss does not correlate with win rate, which has implications for how future Melee models are evaluated and selected during training.
Several natural directions follow from this work. Systematically varying the action delay would clarify its role in evaluation. Extending the curriculum to all 26 characters would test whether character-specific RL fine-tuning is necessary or whether the shared Transformer architecture generalises sufficiently. Applying the same pipeline to other fighting games or real-time strategy games would test the generalisation of the Faynt approach. And studying the emergent character strategies that develop within the shared policy could reveal new insights into Melee's matchup dynamics.