Reinforcement Learning for Out-of-Distribution Reasoning in LLMs: An Empirical Study on Diagnosis-Related Group Coding

Diagnosis-Related Group (DRG) codes are essential for hospital reimbursement and operations but require labor-intensive assignment. Large Language Models (LLMs) struggle with DRG coding due to the out-of-distribution (OOD) nature of the task: pretraining corpora rarely contain private clinical or billing data. We introduce DRG-Sapphire, which uses large-scale reinforcement learning (RL) for automated DRG coding from clinical notes. Built on Qwen2.5-7B and trained with Group Relative Policy Optimization (GRPO) using rule-based rewards, DRG-Sapphire introduces a series of RL enhancements to address domain-specific challenges not seen in previous mathematical tasks. Our model achieves state-of-the-art accuracy on the MIMIC-IV benchmark and generates physician-validated reasoning for DRG assignments, significantly enhancing explainability. Our study further sheds light on broader challenges of applying RL to knowledge-intensive, OOD tasks. We observe that RL performance scales approximately linearly with the logarithm of the number of supervised fine-tuning (SFT) examples, suggesting that RL effectiveness is fundamentally constrained by the domain knowledge encoded in the base model. For OOD tasks like DRG coding, strong RL performance requires sufficient knowledge infusion prior to RL. Consequently, scaling SFT may be more effective and computationally efficient than scaling RL alone for such tasks.

Process-Supervised Reward Models for Clinical Note Generation: A Scalable Approach Guided by Domain Expertise

Process-supervised reward models (PRMs), which verify large language model (LLM) outputs step-by-step, have achieved significant success in mathematical and coding problems. However, their application to other domains remains largely unexplored. In this work, we train a PRM to provide step-level reward signals for clinical notes generated by LLMs from patient-doctor dialogues. Guided by real-world clinician expertise, we carefully designed step definitions for clinical notes and utilized Gemini-Pro 1.5 to automatically generate process supervision data at scale. Our proposed PRM, trained on the LLaMA-3.1 8B instruct model, demonstrated superior performance compared to Gemini-Pro 1.5 and an outcome-supervised reward model (ORM) across two key evaluations: (1) the accuracy of selecting gold-reference samples from error-containing samples, achieving 98.8% (versus 61.3% for ORM and 93.8% for Gemini-Pro 1.5), and (2) the accuracy of selecting physician-preferred notes, achieving 56.2% (compared to 51.2% for ORM and 50.0% for Gemini-Pro 1.5). Additionally, we conducted ablation studies to determine optimal loss functions and data selection strategies, along with physician reader studies to explore predictors of downstream Best-of-N performance. Our promising results suggest the potential of PRMs to extend beyond the clinical domain, offering a scalable and effective solution for diverse generative tasks.