Colon cancer: trained using real-world case data, an 8-billion-parameter Llama model outperforms GPT-5.2 in selecting the initial treatment

Researchers treat therapeutic decision-making in oncology as a detective game: the agent must first gather the right information about the patient before recommending a treatment.

Large language models shine in medical examinations, but clinical decision-making is another thing: a sequential process where information is collected, reasoning with a partial vision of the patient, and where several choices can be defended. This is why learning by clinical reinforcement has so far focused mainly on diagnosis, where a response is more easily verified. A team led by Li-Ching Chen and Julian C. Hong is tackling the more delicate case of front-line treatment of colon cancer.

Rewards from the real journeys

The researchers built an interactive environment in which the officer questioned a structured representation of the patient file before committing to a recommendation. In the absence of a single "good response", rewards are calculated patient-by-patient based on decisions actually made in actual clinical records, assessing information collection, treatment choice and consistency of reasoning.

Optimized with GRPO and GPDO methods, Llama-3.1-8B-Instruct is progressing significantly: the GDPO-trained version improves the average clinical reward from the basic model by 79.6% and exceeds GPT-5.2 by 18.8% in the same environment, while asking fewer questions as the training progresses. The authors see this as evidence that only partially verifiable rewards can guide clinical learning.