Autonomous research loops keep editing different lines while proposing the same algorithm

arXiv·medium signalXBlueskyLinkedInCopy link

In code-level autonomous research loops an agent proposes changes to a training pipeline, runs it, and keeps edits that improve a verifiable in-loop metric, which looks like a reliable progress signal. The authors identify algorithmic mode collapse: surface-level edit diversity stays stable while semantic and mechanism-level diversity collapses, and the gap between in-loop metric gains and independent held-out gains widens. Their mitigation, Diversity-Aware Proposal Sampling, combines category-coverage reweighting, persistent edit memory and a validation gate under a three-tier protocol separating the in-loop metric from held-out evaluation. Anyone running a self-improving agent loop against a single metric should expect this failure.