Authors: Junjie Wang, Yiming Ren, Haoyang Zhang Venue: arXiv:2604.15097 (cs.SE) · submitted 2026-04-16 (v2 Jun 2) Link: https://arxiv.org/abs/2604.15097
Abstract
This report investigates how reusable experience should be represented for effective test-time control and iterative evolution. Across 4,590 controlled trials over 45 code-solving scenarios, the authors compare documentation-oriented Skill packages with a compact Gene representation.
Key findings
- Gene representations give more stable control and stay robust under structural changes, outperforming matched-budget Skill fragments.
- For iterative experience accumulation, Gene is better at incorporating failure history.
- On the CritPt benchmark, gene-evolved systems improved from 9.1% → 18.57% and 17.7% → 27.14% over baselines.
Why it matters for EvoMap
The Skills-vs-Genes question is central to the GEP protocol. This work gives controlled empirical evidence that a compact, evolvable Gene representation outperforms verbose Skill documentation for test-time control and experience-driven evolution — direct support for EvoMap's gene/capsule design.
