MetaSkill-Evolve keeps its base model frozen while editing task skills and five meta-skills that guide diagnosis, retrieval, budget allocation, proposals, and edit execution. The same five-agent pipeline modifies the meta-skills; subsequent iterations load the updated files to perform further improvement. This is an explicit recursive connection, beyond merely editing a task solver. Version 1, §3.4–3.5, Algorithms 1–2.
The default run has five fast iterations and two meta-updates. Evaluation uses held-out partitions of OfficeQA, SealQA, and ALFWorld, not scientific discovery tasks. The pipeline’s roles and wiring remain fixed. Version 1, §4.1, Limitations, and Appendix D.
The control description is inconsistent. Section 4.1’s Single-Level baseline changes branching, sharing, and meta-updates; §4.2 says only meta-updating differs. Appendix E includes a no-meta-updates condition but equates it with that Single-Level row. The text therefore does not clearly establish a comparison changing only recursive feedback under matched resources. We have not independently reproduced the results. Version 1, §4.1–4.3 and Appendix E.
The task-design question
Can an accepted research capability improve the system’s next self-modification?
A scientific campaign could let a revised diagnosis or experiment-selection procedure participate in later improvement work. Compare that feedback with a fixed improver and a condition withholding the revised capability, while matching resources and assessing new scientific problems.
Our classification is a core recursive-mechanism study with evaluation limits. Showing the loop exists does not establish its causal benefit, and these non-scientific benchmarks do not establish scientific transfer.
Explore our draft protocol or help design a scientific task.
Clarification, October 3, 2026: the Brief now traces how updated meta-skills drive later improvement and notes the inconsistent control descriptions, including the no-meta-updates ablation.
Editorial publisher: RSI4Science / DUAL Lab. AI-assisted research, drafting and editing; primary sources checked by automated agents.
Corrections are welcome. Prepare a source-backed correction. Publication does not imply independent experimental reproduction or endorsement of the work discussed.
Turn a research question into a task