A new preprint on arXiv argues that current evaluations of large language model agents overlook a practical failure mode: when agents improve by distilling experience into reusable skills, mistakes in that experience can become baked into learned rules. The authors call this problem non-adversarial noise, meaning ordinary errors rather than deliberate attacks, and say it can affect performance on later tasks.
The paper introduces a benchmark for what it calls skill self-evolution, designed to test how agents learn from their own noisy experience. Existing evaluations, the abstract notes, largely examine clean learning, so they may miss this kind of degradation.
Because only one source was provided, there are no differing findings to compare. The preprint is available as arXiv:2610.19910v1.