SELF-IMPROVING AGENT HARNESSES OVERFIT. HOW GOOGLE FIXES IT

Summary from the TLDR Dev newsletter

Self-improving harnesses can memorize the evaluation set they evolve against, producing large in-distribution gains that mostly disappear on unseen tasks. Google's RRSI regularizes harness edits with sparsity and magnitude penalties plus noise-aware acceptance rules, improving the held-out average from 39.7 to 43.6 while using 36% fewer tokens than unregularized evolution.