The abstract is read at the publisher; the summary is JClub's.
In a tool-first multi-agent biomedical discovery benchmark of 104 single- and multi-turn tests, Gemma 4 (26B) matched larger orchestrators on single-turn performance, while multi-turn evidence and tool-ordering varied.
- Why it matters: Cloud deployment of agentic biomedical discovery raises inference cost, data privacy, and governance concerns, motivating on-premises small-model orchestration that still coordinates scientific tools effectively.
- What they did: The system used an orchestrator coordinating six specialized tool systems (e.g., differential expression, literature retrieval, pathway enrichment, gene prioritization, perturbation matching, structural/druggability), holding tools and prompts constant while comparing Gemma 4 (26B), Gemini 2.5 Flash, and Claude Sonnet 5.
- The result: Single-turn results were comparable with no reviewer ratings of the SLM below larger models; in multi-turn tests, Claude Sonnet 5 had marginally higher evidence validity, Gemma 4 was weakest for cross-turn parameter assembly, and all three struggled with tool ordering on memory-dependent tasks.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.