EviSnap: Faithful Evidence-Cited Explanations for Cold-Start Cross-Domain Recommendation

This document is an arXiv feed (multiple papers) focused on recommender systems, retrieval, and safety/robustness of LLM-based pipelines. Key contributions include: EviSnap — a cold-start cross-domain recommender that produces evidence-cited, faithful rationales by distilling reviews into facet cards and a concept bank; SearchFireSafety — a statute-centric legal-QA benchmark that exposes retrieval gaps and a safety trade-off (domain adaptation improves retrieval but increases hallucination/incorrect confident answers when statutory evidence is missing); an empirical vulnerability report on QWĒ