HAL reportedly reaches roughly top 1% of ranked Super Smash Bros. Melee play

Predicting future inputs from human play data

The article describes HAL as a 247M-parameter causal Transformer trained using human data only. It reads the previous 256 frames of game state and controller inputs, then predicts future controller-input chunks. Its author says it reached roughly top-1% human play on the ranked ladder and can play online in real time under varying hardware and network conditions.

The article does not provide an independently verified, directly comparable benchmark for this model’s claimed ladder standing.

The author reports that scaling pre-training and adding return conditioning plus autoregressive action chunking produced HAL, a 247M-parameter model that can play online in real time while replanning at under 10 Hz.

Responsiveness under online delays

The reported work combines large-scale pre-training with methods intended to make a game-playing model responsive despite inference and network delay. It also reports testing how training scale, model size, observation delay, and action prediction affect play.