Morphogenetic Harness: What LLM Agents Should and Should Not Borrow from Developmental Biology
We introduce the morphogenetic harness, a framework that translates developmental biology principles into LLM agent design, and test which biological concepts actually transfer. Through controlled experiments on standard benchmarks (GSM8K, HumanEval, MMLU) across three LLMs (Claude Haiku, GLM-4-Flash, Grok-3-Mini), we find that static skill routing — inspired by cell differentiation — improves math reasoning by 11-13 percentage points, while evolutionary colony dynamics provide no benefit beyond the pass@k effect. Our key finding: borrow the blueprint, skip the process. The repository includes all experiment code, benchmark data, and statistical analysis.
Authors
- Xinrui Chang
- yuhang xiao
Institutions
- Shanghai Medical College of Fudan University (CN)
- Fudan University (CN)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-04-15
- DOI
- https://doi.org/10.5281/zenodo.19591666
- Primary Topic
- Evolutionary Algorithms and Applications
- Type
- article
- Field-Weighted Citation Impact
- 0.00