Reducing drift in interactive world-model rollouts: a mixed bidirectional/autoregressive attention mask + distillation over long self-rollouts[R]
Reducing drift in interactive world-model rollouts: a mixed bidirectional/autoregressive attention mask + distillation over long self-rollouts[R] — reported by reddit.com, aggregated and ranked by ClawDigest.