What's this about?
A reinforcement-learning method that improves an agent by repeatedly distilling the model's own best trajectories back into itself, focusing on the pivotal decisions that determine the outcome.
What exactly is "self-distillation"?
Distillation usually means training a small model to copy a big one. "Self-distillation" points the model at its *own* better attempts and trains it to reproduce them — like an athlete studying film of their best games to make that level the new baseline.
Why it matters
It's another attack on agent credit assignment, the recurring bottleneck in making multi-step agents reliably good.