Automated fact‑checking is essential for reliable public health information, yet the biomedical domain demands rigorous literature interpretation, evidence assessment, and justification. Existing retrieve‑then‑verify pipelines usually output only a support/refute label, offering little explanatory value for users. To address this, we introduce BioCheck Agent, an LLM‑driven system that produces structured biomedical fact‑checking reports via agentic search. The agent restricts its search to high‑quality PubMed articles using advanced Boolean operators, ensuring domain‑specific accuracy. Because direct prompting often leads to hallucinations, especially for lightweight open‑source models, we propose Evidence‑Grounded Group Relative Policy Optimization (EG‑GRPO), a reinforcement‑learning method that supplies a task‑specific reward encouraging deep search behavior and high‑quality evidence retrieval while penalizing hallucinations. Experiments show that, compared with the base Qwen3.5‑4B model, BioCheck Agent with EG‑GRPO improves label prediction accuracy on SciFact by 9.95%, raises the evidence quality score by 3.7%, and reduces the evidence hallucination rate by 19.63%.
Blogger's Review: This approach leverages reinforcement learning to markedly enhance both retrieval depth and report interpretability, offering a more trustworthy solution for automated biomedical fact‑checking.