Are Rationales Necessary and Sufficient? Tuning LLMs for Explainable Misinformation Detection

In three linesarXiv paper proposing LONSREX, a data synthesis pipeline to fine-tune LLMs for explainable misinformation detection. Authors identify two issues: rationales filtered on binary labels are insufficient, and stronger LLMs produce unnecessarily verbose rationales. LONSREX introduces a metric quantifying the necessity and sufficiency of each verification step.

Summary generated by Claude — human-verified