An annotated reading guide · compiled 6 October 2026 arXiv:2604.15097 · cs.SE
Skill → Gene

Reading notes on From Procedural Skills to Strategy Genes — how LLM agents should encode experience so it actually controls behavior.

arXiv:2604.15097
Wang · Ren · Zhang
Submitted 16 Apr 2026

About · editorial policy

What this guide is, and what it is not

Skill to Gene is an independent annotated reading guide to one paper: arXiv:2604.15097, “From Procedural Skills to Strategy Genes: Towards Experience-Driven Test-Time Evolution” by Junjie Wang, Yiming Ren, and Haoyang Zhang. It exists because the paper’s central result, that a ~230-token Strategy Gene guides a model better than a ~2,500-token documentation package drawn from the same experience, is easy to misquote, and because its tables reward a second pass.

It is not the authors’ site, is not affiliated with them, and does not speak for them. It is not a company page, and nothing is sold here. The paper is the thing under review; this guide is marginalia around it.

Who edits this guide

The byline on every page names the paper’s three authors — Junjie Wang, Yiming Ren, and Haoyang Zhang — credited as the paper’s authors, because their names belong at the top of anything written about their work. The annotation itself carries a separate, collective credit, “the Skill to Gene editors,” which is the accurate description: the guide is maintained by a small editorial desk of technical readers who follow the agent-memory literature, and no individual editor byline is claimed. The desk is deliberately easy to check up on. It is not affiliated with the paper’s authors, publishes under a single name, sells nothing, and makes its sourcing rules explicit so that any reader can re-verify an annotation without trusting anyone. Where the guide’s competence ends — transfer beyond scientific code-solving, the correctness of the paper’s underlying experiments — it says so instead of speculating.

How this guide was made

The working source is the paper itself: the arXiv abstract page and the HTML rendering of v1, read closely and kept open beside every drafting session. Drafts of each page are prepared with AI assistance, then verified by an editor line by line — every figure traced back to its source table, every quotation checked character-for-character, every claim that comes from outside the paper tagged as interpretation. That verification pass is the substance of the guide; the drafting tools are not. The process runs in one direction: when a check and a draft disagree, the draft loses.

Two consequences follow. First, nothing on this site is quoted from another website’s summary of the paper — if you see a number here, it was read off the paper’s own abstract, tables, or appendices. Second, the guide gets revised against the source rather than defended: when arXiv posts a revision, the tables here are re-checked before the site claims to still describe them.

Sourcing rules

  • Every number on this site is quoted from the paper’s abstract, tables, or appendices, with the source table named where one exists.
  • Where the paper reports a sub-experiment with its own baseline, we reproduce that baseline rather than blending runs — the failure-ordering table on the findings page is the standing example.
  • Interpretations that are not in the paper (the “life paradigm” reading, the compiled-artifact analogy) are attributed to the commentary that makes them and labeled as interpretation.
  • Caveats travel with their claims: benchmark scope, model specificity, and the beta status of the report are repeated in place, not footnoted away.

Corrections

If a quote is wrong, a table drifted from its source, or a caveat is missing, write to [email protected] and it will be fixed against the paper. Changes are made to match the source, not to defend the annotation.