EvoMap Publishes Skill to Gene Research Showing How AI Agents Can Keep Learning Across Runs
EvoMap, the open infrastructure for AI self-evolution, outlines how its Genome Evolution Protocol (GEP) reflects arXiv:2604.15097. The report found that compact Strategy Genes guide AI agents better than documentation-style Skill packages.
United States, 14th Oct 2026 — EvoMap, which builds open infrastructure for AI self-evolution, has outlined how its Gene and Capsule model addresses a question agent builders keep returning to: what should an agent carry from one run into the next?
Most agent stacks answer by saving more: skill files, tool documentation and retrieved memory. A growing number of builders now see the real issue as what to keep, not how much, a view captured in one recent analysis of the shift from skill to gene.
What the Report Found
That question is at the center of a beta technical report, “From Procedural Skills to Strategy Genes: Towards Experience-Driven Test-Time Evolution” (arXiv:2604.15097), by Junjie Wang, Yiming Ren and Haoyang Zhang, with Infinite Evolution Lab, EvoMap among the listed affiliations. Across 4,590 controlled trials in 45 scientific code-solving scenarios, a compact Gene representation averaged 54.0%, against 49.9% for a full documentation-style Skill package.
A one-line lesson can get lost in the documentation that surrounds it, while a Gene keeps it in front of the model. Failure history likewise helped most when distilled into short warnings rather than appended wholesale. In the report’s CritPt runs, gene-evolved systems improved on their paired base models from 9.1% to 18.57% and from 17.7% to 27.14%.
From Research to Practice
EvoMap’s Genome Evolution Protocol (GEP) applies a similar structure. A Gene is a reusable strategy, a Capsule is a validated fix with an audit trail, and agents can share, verify and inherit both. Experience stays editable as new evidence arrives instead of piling up as an ever-longer prompt, a shift EvoMap’s own blog describes as moving from static tools to evolving Genes.
“An agent’s experience only counts if the next run can pick it up and act on it,” said Haoyang Zhang, Founder of EvoMap. “Turning what an agent has learned into a compact, reusable strategy is the common thread linking the research question and the protocol.”
Developers building on open-source agent frameworks have picked up the idea too. One community article frames it as compiling experience rather than merely storing it, and one practitioner sums up the rule as “save the lesson, not the transcript.” In Japanese, one analysis of the paper separates memory, which records what happened, from a gene, which sets a rule for what to do next. Several write-ups have circulated under the headline “From Skill to Gene: Why AI Agents Need to Evolve from the Tool Paradigm to the Life Paradigm.”
About EvoMap
EvoMap is an open infrastructure for AI self-evolution. Through its Genome Evolution Protocol, AI agents can share, validate and inherit reusable capabilities, organized as Genes (reusable strategies) and Capsules (validated fixes with audit trails). Learn more at https://evomap.ai/.
Media Contact
Organization: evomap ai
Contact Person: Haoyang Zhang
Website: https://evomap.ai/
Email: Send Email
Country: United States
Release id: 49841
View source version on King Newswire:
EvoMap Publishes Skill to Gene Research Showing How AI Agents Can Keep Learning Across Runs