AI agents that retrain themselves on their own output tend to get worse, not better, after a few rounds - a new method claims to fix that.
Researchers studying "iterative self-distillation," where an AI agent trains on its own past performance and the resulting model becomes the teacher for the next round, found that existing methods degrade over time instead of improving. Success rates fall with each cycle, and so does the model's ability to use privileged information, extra context fed in during training that isn't available at deployment. The paper's fix, Retentive and Selective Augmentation for Iterative Self-Distillation (ReSAIL), targets the training steps where that privileged information most changes the model's predictions and balances them across the data, while anchoring the student's behavior to a frozen copy of the teacher so later rounds don't erase what earlier ones learned. On the ALFWorld and TextCraft agent benchmarks, adding ReSAIL to existing self-distillation methods produced an average 22.5% absolute gain in success rates after three training cycles, and a related data-selection technique also improved action prediction for GUI agents on a benchmark called AITZ.
This matters because "recursive self-improvement," models training the next generation of themselves, is a mechanism several frontier labs are quietly counting on, and a collapse-prone process undercuts that whole bet. ReSAIL doesn't make agents smarter in any absolute sense; it just stops a specific failure mode where repeated self-training erases earlier progress, which is a narrower and more useful claim than the framing suggests.
That's a modest, well-scoped result dressed in a maximalist acronym - worth watching, but it's an engineering patch, not evidence that agents are on a path to improving themselves without limit.