rollout-training

This paper proposes HERO (History-Enriched Rollout Training), a method that augments standard trajectory supervision for autoregressive neural operators with relative supervision from the model's own optimization history, improving long-horizon accuracy and stability on PDE benchmarks.