🤖 AI Summary
This study addresses the challenge of defining reinforcement learning rewards in clinical question answering, where executable verifiers are unavailable and subtle entity-level errors are prevalent. To this end, it proposes a soft verification method grounded in concept overlap within the UMLS ontology. This approach pioneers the use of controlled-vocabulary concept overlap as a graded external reward signal, integrating an entropy-normalized LLM judge with consistency penalties and optimizing model training via the GRPO algorithm to effectively capture entity substitution errors and enhance safety evaluation. Experimental results demonstrate that the proposed model achieves up to a 39% improvement in Token-F1 on MedQA and PubMedQA, significantly outperforming supervised fine-tuning baselines while exhibiting strong transferability.
📝 Abstract
Reinforcement learning post-training for language models relies on two reward designs: human preferences (RLHF, DPO) and binary verifiers (RLVR). Clinical question answering fits neither. Near-correct answers differ by a single substituted entity, and no executable check decides clinical correctness. We instantiate a soft verifier from a maintained controlled vocabulary: UMLS Concept Unique Identifier overlap (via scispaCy, set-level F1) gives a graded, externally specified reward computed without a model in the loop. We combine it inside GRPO with an entropy-normalised LLM judge, which covers the safety and evidence axes overlap cannot see, and a small consistency penalty on padding and repetition that keeps early-training samples scorable. This three-term composite improves over SFT on Phi-3-mini (3.8B) over MedQA by 2.9% on EM (0.700 vs 0.680) and 39% on Token-F1 (0.202 vs 0.145); on Llama-3.2-3B the corresponding gains are 14% on EM and 35% on Token-F1. We report Token-F1 as the primary metric because it credits partially-correct clinical content that EM discards at this open-generation scale. Main-table results are means over 3 seeds with standard deviations below 0.005. The method transfers to PubMedQA, where training on the PubMedQA train set with the same composite reward improves Token-F1 over SFT by 22% on Phi-3-mini and 17% on Llama-3.2-3B without retuning. A reward ablation on Phi-3, varying the judge-ontology split at a fixed consistency weight, attributes 3 EM points to the ontology term, the contribution that catches entity substitutions the judge cannot. Three negative findings constrain the design: DPO under random negatives underperforms SFT for strong-prior models but helps the weakest-prior one; PPO under a sparse neural reward diverges; GRPO with KL-in-loss collapses at 7B.