AI RESEARCH
CARO: Chain-of-Analogy Reasoning Optimization for Robust Content Moderation
arXiv CS.AI
•
ArXi:2604.10504v1 Announce Type: new Current large language models (LLMs), even those explicitly trained for reasoning, often struggle with ambiguous content moderation cases due to misleading "decision shortcuts" embedded in context. Inspired by cognitive psychology insights into expert moderation, we