Does Feedback Format Change LLM-Guided Biological Search?

Large language model agents can select experiments over sequential biological rounds, but studies disagree about how strongly their choices depend on observed outcomes. We ask whether adding redundant row-level outcome labels changes feedback use in closed-loop biological search. We fix the model, task instructions, assay budget, and biological data, and run a 12-replicate pilot on a genome-wide CRISPR knockout screen using a Sonnet 5 route through NVIDIA Inference. Each replicate compares two presentations, grouped headings alone or the same headings plus row labels, with three feedback conditions: correct outcomes, outcome pairs shuffled among genes within each batch, or outcomes withheld. All six campaigns share the same first-round proposal. The primary outcome is the number of distinct true-hit genes identified per 100 assay slots after feedback begins. Correct feedback yielded more hits than shuffled feedback under both presentations, with an advantage of 2.69 hits per 100 assay slots under grouped headings and 1.48 under row labels. Contrary to our hypothesis, the difference between these advantages was $-1.20$ (95\% interval $[-2.93, 0.54]$; $p = 0.230$). Correct feedback also yielded more hits than withholding feedback altogether. Since shuffling preserved the overall batch yield while breaking the gene-outcome mapping, these results suggest that the agent used gene–specific outcome information to guide its search. However, we found no evidence that row labels strengthened this effect.