LegalHalBench
A benchmark for evaluating hallucination in legal AI systems, primarily assessing statutory relevance. Its main limitation is the difficulty of capturing complete grounding relationship chains between legal claims and their authoritative sources.
Timeline (last 90 days)
The paper identifies limitations of existing evaluation paradigms by name, including answer accuracy, citation existence, general attribution, LegalHalBench-style statutory relevance, and CitaLaw-style sentence-citation alignment.
LegalHalBench primarily tests model recognition ability by constructing questions containing fabricated statutes, focusing on detecting whether non-existent rules have been cited.
All Facts (2)
The paper identifies limitations of existing evaluation paradigms by name, including answer accuracy, citation existence, general attribution, LegalHalBench-style statutory relevance, and CitaLaw-style sentence-citation alignment.
50%UnverifiedLegalHalBench primarily tests model recognition ability by constructing questions containing fabricated statutes, focusing on detecting whether non-existent rules have been cited.
50%