GRPO with variance-aware rubric rewards boosts heart-focused medical QA

Arxiv CS.CL (NLP/language models)

The paper introduces variance-aware rubric rewards with GRPO to improve LLM accuracy on cardiology-related medical questions, achieving significant gains over standard supervised fine-tuning. The method addresses both answer correctness and confidence calibration without requiring additional annotated data.

llm

Arxiv CS.CL (NLP/language models)·Arash AhmadiParisa MasnadiSarah SharifCharles NicholsonDavid EbertMike Banad

6 days ago