EvoSkill v2 preprint reports gains from agents rewriting task instructions

An EvoSkill v2 preprint reports that agents improved held-out task performance by rewriting their instructions while leaving the underlying model and harness unchanged. With fine-grained ground truth, the system exceeded the stronger of SkillOpt and SkillBoost by 7.7 percentage points on SpreadsheetBench v2 and 5.4 percentage points on a banking benchmark. A persistent coding agent reviews training attempts and revises instructions each epoch. Across both benchmarks and feedback settings, its optimizer cost less per percentage point of held-out improvement than either baseline. The results concern benchmark tasks graded with supplied feedback.