Arcee trained the open Trinity Mini model in two scientific environments: Drug Tool for using research tools and BioReason for biomedical reasoning. The team reports 21 controlled runs using GRPO and LoRA, separated training and held-out sets, and deterministic accounting for tool calls.
In the selected run, Arcee says Drug Tool performance rose from 70.8 to 81.2 and BioReason reached 0.863. Those results come from the authors' own environments, but the article exposes enough of the training and evaluation structure to inspect where the gain came from.
The reusable lesson is that a scientific agent is not created by a domain prompt alone. The task distribution, available tools, reward, and verification harness jointly define what the model learns. Publishing those boundaries makes the result reproducible and gives critics something more useful than a single benchmark score.