Self-evolving agents hit a capability-contamination cliff: skill pools degrade past a critical size, and rollback doesn't fix it

arXiv·high signalXBlueskyLinkedInCopy link

Shang et al. (arXiv 2608.05810, submitted August 6) show that agents distilling reusable skills from their own trajectories improve only up to a critical pool size, after which new skills actively degrade performance — a defective skill becomes reference material for distilling later skills, forming cross-round contamination chains. The damage is structurally irreversible: deleting the source skill afterward recovers only a small fraction of the lost performance, because descendants already inherited the flawed reasoning. Their Verifier-as-Gatekeeper admits skills through three non-substitutable critics (structural validity, behavioral harmlessness, semantic consistency) plus marginal-gain subset selection, improving every round to 72% pass@1 on Terminal-Bench 2 with a pool roughly 5x smaller, and the frozen pool transfers to four other backbones without re-evolution.