Toward a Locally Deployable Agentic Co-Scientist: Small-Model Planning for Early-Stage Drug Discovery
A compact local model can plan drug-discovery tool workflows, but still struggles when the full workflow path is new.
The paper describes a lightweight agent framework where a small language model routes requests across 18 modular scientific tools. Its shared molecular schema keeps records consistent, and the authors fine-tune compact models on 1,263 curated query-plan pairs. On held-out queries with familiar workflow patterns, Llama 3.2-3B nearly matched the target plans; on stricter unseen workflow sequences, exact-match performance fell to roughly 45–55%. ArXiv · AI/CL/LG's note
The paper describes a lightweight agent framework where a small language model routes requests across 18 modular scientific tools. Its shared molecular schema keeps records consistent, and the authors fine-tune compact models on 1,263 curated query-plan pairs. On held-out queries with familiar workflow patterns, Llama 3.2-3B nearly matched the target plans; on stricter unseen workflow sequences, exact-match performance fell to roughly 45–55%. ArXiv · AI/CL/LG's note
score 4