Case Study: Loka achieves verifiable scientific reasoning with Arcee AI’s Trinity Mini

Loka boosts held-out Drug Tool accuracy to 81.2% with Arcee AI

Loka, a biomedical research company, collaborated with Arcee AI to overcome the challenge of teaching an AI model to perform reliable, evidence-based scientific research. Their goal was to move beyond simple classification and enable a model to investigate complex biological questions using various tools, reason across incomplete data, and produce auditable, structured results that another researcher could verify.

Arcee AI implemented a solution using their Trinity Mini open model, which was post-trained with a specialized LoRA adapter. They used a reinforcement learning process with two custom environments to teach the model investigative and reasoning skills. The result was a highly specialized model, known as run 120, which achieved 81.2% accuracy on a held-out drug-tool evaluation and a composite score of 0.863 on a biological reasoning task. This allowed Loka to deploy the adapter within a secure, controlled AI Scientist application for practical, tool-assisted biomedical research.

Loka