Title: Enhancing Clinical Summarization through Hallucination Detection-Informed Preference Optimization

Abstract:

While Large Language Models (LLMs) have demonstrated significant potential in summarization tasks, their utility in specialized healthcare contexts is often compromised by hallucinations—unsupported or inaccurate statements that undermine reliability. To address this challenge, we present \itermodelfull (\itermodel), an inference-time technique that utilizes hallucination detectors to steer iterative revisions of summaries toward factual accuracy. Building upon this foundation, we introduce \itermodel for Preference Learning (\model), a method that transforms these detector-guided refinement trajectories into preference pairs for model fine-tuning.

Our extensive experiments, conducted using real-world clinical notes from \MimicIV, reveal that these approaches significantly mitigate hallucinations in both Llama and Gemma models. Specifically, \itermodel achieves a 24\% reduction in hallucinations, while \model yields an even more substantial 48\% decrease for the Llama-3.1-8B-Instruct model. Crucially, evaluations by human experts and an LLM-Jury confirm that both methods maintain high standards of fluency, coherence, and relevance in the generated summaries. Collectively, these findings highlight that combining detection-informed refinement with preference learning provides an effective, automated strategy for enhancing factual faithfulness in clinical text summarization.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC