Back to feed
arXiv cs.CL·

Hallucination Detection-Guided Preference Optimization for Clinical Summarization

Signal
78
Hype
15
In three linesPreference optimization method guided by hallucination detection to improve clinical summarization reliability. On Llama-3.1-8B-Instruct, reduces hallucinations by 24% at inference and 48% after fine-tuning, preserving fluency. Evaluated on MIMIC-IV.
Read source
Your take?
LlamaFine-tuningAI safetyEvalsPapers

Summary generated by Claude — human-verified