Hallucination Detection-Guided Preference Optimization for Clinical Summarization

Researchers introduce HDPO, a method that uses hallucination detectors to guide iterative refinement of AI-generated clinical summaries, reducing factual errors by up to 48% in large language models. The approach combines inference-time detection with preference learning for model finetuning, demonstrating significant improvements in factual accuracy while maintaining summary quality for healthcare applications.

This research addresses a critical vulnerability in deploying large language models within healthcare systems—their tendency to generate hallucinations that could compromise clinical decision-making. The work presents two complementary approaches: HDPO operates at inference time by iteratively refining summaries based on hallucination detection signals, while HDPO-PL converts these refinement trajectories into training preference pairs for permanent model improvement. The 48% hallucination reduction in Llama-3.1-8B represents substantial progress toward clinical-grade reliability.

The healthcare sector has increasingly adopted LLMs for documentation and summarization tasks, where errors carry serious consequences. Traditional safety approaches focus on pre-training guardrails or post-hoc filtering, but this detection-guided methodology embeds validation directly into the generation process. The use of real-world clinical notes from MIMIC-IV grounds findings in authentic complexity rather than synthetic benchmarks.

For healthcare AI developers and medical institutions, this methodology reduces deployment risk by addressing hallucinations without sacrificing fluency or clinical relevance—metrics verified through expert human evaluation and LLM-based assessment. The preference learning component suggests a path toward automated dataset curation for fine-tuning, potentially lowering the cost of specialized model adaptation. Healthcare providers evaluating LLM investments gain a practical framework for improving model reliability, though implementation requires access to hallucination detection models and sufficient computational resources for iterative refinement.

• →HDPO-PL reduces hallucinations by 48% in Llama-3.1-8B while preserving summary quality and clinical relevance.
• →Detection-guided preference learning enables automated conversion of refinement trajectories into effective training data.
• →The method operates on real-world clinical documentation, demonstrating practical applicability beyond synthetic benchmarks.
• →Both inference-time and finetuning approaches preserve fluency and coherence according to expert evaluations.
• →The framework provides healthcare institutions a scalable pathway to improve LLM reliability for clinical applications.