FaithMed: Training LLMs For Faithful Evidence-Based Medical Reasoning

Medical LLMs often struggle to maintain faithfulness, frequently generating plausible-sounding rationales that are not actually supported by retrieved evidence. This paper investigates whether explicit supervision of the reasoning process—rather than just the final answer—can improve the reliability and evidence-grounding of medical LLMs.

To address this, the authors introduce FaithMed, a framework that formalizes evidence-based medicine (EBM) principles into five process-level dimensions: Ask, Acquire, Appraise, Apply, and Assess. The researchers collaborated with clinicians to design rubrics for these dimensions, which were then automatically refined to ensure they were discriminative and non-redundant. FaithMed uses these rubrics to assign step-level process rewards during reinforcement learning. By grouping reasoning steps with similar contexts, the model learns to compare its intermediate actions against others, effectively training it to perform clinically sound reasoning steps like formulating precise search queries and critically evaluating retrieved evidence.

FaithMed significantly outperforms standard agentic-search and outcome-only reinforcement learning baselines across seven medical benchmarks. On average, it improves answer accuracy by 9% over agentic-search methods and achieves a 15.5% gain in EBM rubric scores. The authors demonstrate that step-level process rewards are crucial; they provide more consistent, fine-grained feedback than episode-level rewards, which often suffer from noisy credit assignment. Furthermore, ablation studies confirm that the rubric dimensions target distinct reasoning stages, and that improvements in these intermediate steps directly correlate with higher final task performance.

This work shifts the focus of medical LLM training from simple output correctness to the faithfulness of the underlying reasoning process. By demonstrating that models can be trained to follow EBM workflows, the authors provide a scalable path toward more transparent and trustworthy clinical AI assistants that justify their conclusions through verifiable evidence.