Title: From Procedural Skills to Strategy Genes: Towards Experience-Driven Test-Time Evolution

Abstract:

This technical report, currently in beta, investigates the optimal representation of reusable experience to serve as both effective test-time control and a foundation for iterative evolution. To address this, we conducted 4,590 controlled trials spanning 45 distinct scientific code-solving scenarios. Our analysis reveals that documentation-oriented Skill packages offer inconsistent control; their actionable signals are sparse, and inflating a concise experience object into a comprehensive documentation package typically fails to provide benefits, often reducing overall performance averages.

Furthermore, we demonstrate that the method of representation is a primary determinant of success. A compact "Gene" representation achieves the highest overall average, maintains competitiveness even under significant structural perturbations, and surpasses Skill fragments with equivalent budgets. Conversely, adding documentation-oriented content to Genes generally diminishes rather than enhances their utility.

Beyond one-shot control, our findings indicate that Genes are superior carriers for accumulating experience iteratively. Failure history integrated into Genes proves more effective than when attached to Skills or freeform text. Additionally, the editable structure of the experience object is critical, with failure information yielding the greatest benefit when distilled into concise warnings rather than being naively appended. On the CritPt benchmark, gene-evolved systems demonstrated improvements over their paired base models, rising from 9.1% to 18.57% and from 17.7% to 27.14%. These outcomes suggest that the fundamental challenge in experience reuse lies not in providing greater volumes of data, but in encoding experience as a compact, control-oriented object ready for evolution.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC