Biological Context Does Not Reliably Improve LLM-Guided Experimental Optimization
Researchers tested whether providing LLMs with biological context improves their ability to optimize experimental designs compared to ignoring that context. They found that for biology tasks, domain-aware prompting actually made models perform worse than domain-agnostic prompting on 58.7% of tasks, suggesting biological knowledge does not reliably help LLMs choose better designs in this setting.