A recent paper published on arXiv cs.CL investigates the relationship between answer correctness and the citation of legal authority in large language model (LLM) responses to legal questions. The research tested four LLMs using 238 Taiwan bar-examination items, which included verified governing provisions. The study aimed to determine if scoring final answers adequately reflects a model's grounding in legal authority.
Key Points
- Legal benchmarks typically score final answers, even when models provide legal authority. This study tested if answer correctness serves as a proxy for authority grounding.
- Four LLMs spontaneously produced authority markers for 238 Taiwan bar-examination items, even when not explicitly prompted for statutory citations.
- The study jointly audited answer correctness and authority grounding, finding that these two dimensions can dissociate.
- In criminal law, 24.0% to 42.4% of valid responses were correct in their answer but missed the gold authority.
- Conversely, 15.2% to 21.7% of responses were incorrect in their answer but correctly cited the gold authority.
- A separate statutory-retrieval probe and a citation-abstention intervention further demonstrated that answer and citation behaviors can diverge at the output level.
- This mismatch occurred without adversarial or inconsistency-inducing prompting.
Context
According to the authors, legal benchmarks often focus solely on the correctness of the final answer, overlooking the model's ability to cite relevant legal authority. This approach assumes that a correct answer implies proper grounding in legal statutes. The study's methodology involved using Taiwan bar-examination items, which provide a clear gold standard for both the correct answer and the governing legal provision. This allowed for an automatic audit of both dimensions simultaneously, revealing instances where models produced correct answers without citing the correct authority, and vice-versa.
Why It Matters
This research highlights a critical limitation in current legal benchmark evaluation practices. If answer-only scoring treats naturally occurring misses of gold authority as complete successes, it may lead to an overestimation of an LLM's true legal reasoning capabilities. For builders, this suggests that models might perform well on answer correctness metrics without genuinely understanding or applying the underlying legal principles, posing risks in applications requiring precise legal grounding.
What To Do
- Note that answer correctness alone may not indicate proper authority grounding in legal applications.
- Consider incorporating authority grounding as a distinct evaluation metric in legal benchmarks.
- Review the paper's methodology for automatically auditing answer correctness and authority grounding.
- Watch for similar decoupling behaviors in other domain-specific benchmarks where external verifiable sources are critical.