PAPER / ARXIV:2609.09754
Yujin Zhou, Mingxuan Zheng, Chuxue Cao, Huang Yidan, Jiale Chen, Yike Guo, Sirui Han
RESUMO
Tool-augmented legal agents introduce agentic hallucinations where tool-call and reasoning errors cascade into fabricated holdings and miscited authority. We introduce LexAgentHallu, containing 3414 instances across 17 legal categories, annotated under a dual-layer hallucination taxonomy of 7 categories and 27 subclasses. Evaluation across 18 agents uncovers a Right-Answer-Wrong-Reason effect and shows hallucination subclasses cluster into distinct agentic-framework and legal-task profiles, invisible to outcome-level evaluation.
NO MESMO MAPA