ReSAIL: Mitigating Collapse in Iterative Agent Self-Distillation
ReSAIL is meant to stop agent self-distillation from degrading over repeated deployment cycles.
The paper says existing iterative self-distillation methods lose deployment performance over cycles, even when privileged-information behavior is involved. ReSAIL selects interaction steps where privileged information most changes the teacher’s predictions, then balances the distillation loss across trajectories. It also regularizes the student against a frozen teacher so that privileged-information-conditioned behavior survives into the next teaching cycle. On ALFWorld and TextCraft, adding ReSAIL produced an average 22.5-point final-cycle success-rate gain over self-distillation baselines. Source: HF Daily Papers' note.
The paper says existing iterative self-distillation methods lose deployment performance over cycles, even when privileged-information behavior is involved. ReSAIL selects interaction steps where privileged information most changes the teacher’s predictions, then balances the distillation loss across trajectories. It also regularizes the student against a frozen teacher so that privileged-information-conditioned behavior survives into the next teaching cycle. On ALFWorld and TextCraft, adding ReSAIL produced an average 22.5-point final-cycle success-rate gain over self-distillation baselines. Source: HF Daily Papers' note.
score 5