AI Agent Improves Biomedical Fact-Checking with RL and Agentic Search

A new LLM-based agent, BioCheck Agent, has been introduced to automate the generation of structured fact-checking reports specifically for the biomedical domain. Unlike existing methods that merely output a "supported" or "refuted" label, BioCheck Agent performs agentic searches exclusively on high-quality scientific literature in PubMed, then synthesizes a comprehensive conclusion with supporting evidence and detailed analysis. To overcome common LLM issues like hallucinations and low-quality reports, especially with smaller open-source models, the researchers implemented Evidence-Grounded Group Relative Policy Optimization (EG-GRPO). This reinforcement learning technique incentivizes advanced search behaviors and high-quality evidence retrieval while penalizing inaccuracies. Experimental results show significant improvements in label prediction accuracy, evidence quality, and a substantial reduction in hallucination rates compared to baseline models, demonstrating its ability to produce more reliable biomedical fact-checking reports.