What's this about?
ReSAIL selects the interaction steps where hidden teacher information matters most and preserves the teacher's informed output distribution. The authors report a 22.5-point average final-cycle gain over self-distillation baselines.
What exactly is self-distillation?
An agent uses traces from an earlier, better-informed version of itself as training data for the next version. Repeating this can compound useful experience—or compound errors until performance collapses.
Why does this matter?
Learning repeatedly from deployment is central to continual agents, so preventing generational decay is important. Evidence currently covers three cycles and selected environments, not indefinite recursive improvement.