Reproduce “MedFact-R1: Towards Factual Medical Reasoning via Pseudo-Label Augmentation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Gengliang Li, Rongyu Chen, Bin Li, Linlin Yang, Guodong Ding
MEDFACT-R1 introduces a two-stage framework that combines pseudo-label generation with Group Relative Policy Optimization (GRPO) and tailored reward signals to enhance factual medical reasoning in Vision-Language Models. The system achieved up to 22.5% absolute improvement in factual accuracy over previous state-of-the-art methods across multiple medical QA benchmarks.
Key insights
- Pseudo-label SFT effectively addresses medical data scarcity, providing a strong initial knowledge base for VLMs that significantly improves performance and reduces hallucinations.
- The specialized GRPO stage, driven by tailored factual rewards, further refines the model's reasoning, encouraging structured, factually grounded, and logically consistent outputs.
- The combined two-stage approach enables smaller models to surpass larger SOTA models, demonstrating the efficiency and effectiveness of integrating knowledge grounding with advanced RL.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.