LEMON: Equivariant Learning over Attention Maps for Zero-Shot Misbehavior Monitoring on New LLMs

Recent research introduces LEMON, a method for zero-shot misbehavior monitoring in LLMs using equivariant learning over attention maps to reduce expensive re-annotation costs. Simultaneously, a new framework called TABHGR+ enhances zero-shot tabular learning. Unlike traditional row-wise prompting, TABHGR+ utilizes LLM-generated rules to label and group rows via hypergraph reasoning. This approach improves efficiency and better captures inter-row dependencies in diverse domains like finance and healthcare, overcoming high inference costs associated with current LLM methods.