💬 NLP

Knowing the Form, Not the Function: Automatically Auditing Answer--Authority Decoupling in Legal Benchmarks

This paper proposes a joint evaluation framework for legal benchmarks that simultaneously assesses answer correctness and statutory authority grounding, revealing that models frequently decouple these dimensions and that relying solely on answer accuracy leads to misleading performance metrics.

Original authors: Hsien-Jyh Liao

Published 2026-08-05

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Hsien-Jyh Liao

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are in a giant, high-tech library where robots are learning to be lawyers. In this world, there's a special game called a "benchmark." It's like a final exam where a robot reads a tricky legal question and picks the right answer. For a long time, the teachers (the scientists) only cared if the robot got the final answer right. If the robot said "The correct choice is B," it got a gold star. But here's the catch: sometimes, the robot might say "B" but give a completely made-up or wrong reason for why it's B, like citing a law that doesn't exist or pointing to the wrong rulebook.

This paper dives into a specific corner of that library: the world of Legal Benchmarks. Think of these benchmarks as the standardized tests for AI lawyers. The key idea the paper explores is Authority Grounding. In plain English, this means checking if the robot's answer is actually supported by the real, official laws it claims to be using. It's the difference between a student guessing the right answer on a math test versus actually showing the correct formula. The author cares about this because if we only grade the final answer, we might think a robot is a brilliant lawyer when it's actually just a lucky guesser who happens to sound confident.

The researchers decided to play a game with four different AI models (think of them as four different robot students) using real questions from the Taiwan Bar Exam. They asked the robots to solve legal puzzles and explain their reasoning, but they didn't explicitly tell them, "Hey, you must quote the specific law article." Surprisingly, the robots started quoting laws on their own, just like a real lawyer would.

Here is the big surprise the paper found: The robots often got the answer right but the law wrong, and sometimes they got the answer wrong but the law right. It's like a student getting the right answer on a history test but citing the wrong century as their source. The paper calls this a "double dissociation," which is a fancy way of saying the two skills—getting the answer right and finding the right law—can completely separate from each other.

In the criminal law section of the exam, the paper measured exactly how often this happened. They found that between 24.0% and 42.4% of the time, a robot gave a correct answer but missed the correct law entirely. Even stranger, between 15.2% and 21.7% of the time, a robot gave the wrong answer but correctly cited the right law. This proves that just because a robot gets the final answer right, it doesn't mean it actually understands the legal rules behind it.

The author also tried a little experiment. They told the robots, "If you aren't sure of the exact law number, you don't have to say it." When they did this, the robots stopped quoting laws as much, but their ability to get the final answer right barely changed. This suggests that the robots weren't really using the laws to figure out the answer; they were just adding the laws in as decoration after they had already guessed the answer.

So, what's the takeaway? The paper argues that we need to change how we grade these AI lawyers. We can't just look at the final answer anymore. We have to check the "receipts"—the specific laws they cite. If we don't, we might be celebrating a robot that is actually just hallucinating legal facts while getting lucky with the right answer. The author suggests that future tests should score both the answer and the law citation together, because right now, the current system is missing a huge chunk of the truth.

Technical Summary: Knowing the Form, Not the Function

Problem Statement
Current legal benchmarks for Large Language Models (LLMs) typically evaluate performance based solely on the correctness of the final answer, even when the model's reasoning includes citations to legal authorities. This "answer-only" scoring implicitly assumes that a correct final answer implies that the asserted legal authority (the "gold authority") is also correct. The paper argues that this assumption is flawed: a model may arrive at the correct conclusion through valid reasoning or elimination while citing an incorrect or non-existent statutory provision. Because statutory authority is structurally extractable and externally verifiable, the failure to ground an answer in the correct law constitutes a measurable validity problem that current benchmarks systematically hide.

Methodology
The author proposes a joint evaluation framework that decouples "answer correctness" ( $A$ ) from "gold-authority grounding" ( $G$ ).

Testbed: The study utilizes 238 items from the 2022 and 2025 Taiwan National Bar Examinations (118 civil law, 120 criminal law). Each item has a verified correct answer and a verified governing statutory provision (gold authority), derived from commercial solution manuals and cross-checked against the official Laws and Regulations Database.
Models: Four LLMs were evaluated: Gemma-4-31B, LLaMA-3.3-70B, GPT-4o-mini, and Qwen3.6 Flash.
Experimental Conditions:
- C1 (Baseline): Answer-only prompts (no reasoning required).
- C2 (Core): Standard reasoning prompts where models are asked to provide step-by-step reasoning but are not explicitly asked to cite statutes. This tests for "spontaneous" authority marking.
- C3 (Retrieval Probe): Models are explicitly asked to identify up to three governing provisions to measure retrieval capability independent of the answer.
- C4 (Epistemic Probe): Models are asked to recite specific statute content to test uncertainty admission.
- C2-1 (Intervention): A permissive prompt allowing models to reason without naming exact article numbers if unsure, to test if citation behavior can be decoupled from answer selection.
Metrics: Responses are classified into a joint state $(A, G)$ $(A, G)$ :
- $A^+G^+$ : Correct answer, correct authority.
- $A^+G^-$ : Correct answer, incorrect/missing authority.
- $A^-G^+$ : Incorrect answer, correct authority.
- $A^-G^-$ : Incorrect answer, incorrect/missing authority.
- Authority matching is rule-based, parsing code names, article numbers, and paragraph suffixes.

Key Findings
The study identifies a systematic, non-induced "double dissociation" between answer correctness and authority grounding across all tested models:

Answer-Correct but Authority-Missing ( $A^+G^-$ ): A significant portion of responses with correct answers fail to cite the gold authority.
- In Civil Law, this ranged from 32.7% to 63.1%.
- In Criminal Law, this ranged from 24.0% to 42.4%.
- Crucially, these models often cited non-gold statutes that were syntactically valid but legally irrelevant to the specific question.
Answer-Incorrect but Authority-Present ( $A^-G^+$ ): Models frequently cited the correct governing statute even when selecting the wrong answer option.
- In Criminal Law, this occurred in 15.2% to 21.7% of valid responses.
Independence of Behaviors:
- Retrieval vs. Grounding: Independent statutory retrieval (C3) varied significantly by model and domain (e.g., Qwen showed high retrieval rates in civil law), yet the mismatch between answer and authority remained substantial even in models with better retrieval.
- Citation vs. Answer: The C2-1 intervention demonstrated that citation behavior is separable from answer selection. When allowed to abstain from naming specific articles, citation rates dropped dramatically for Gemma, LLaMA, and GPT-4o-mini (from ~70–90% to ~5%), while Qwen's rate dropped from 96.4% to 57.3%. In all cases, answer accuracy remained largely unchanged.
Form over Function: Models frequently produced "fluent" legal reasoning with syntactically correct citations that did not support the conclusion. The presence of a citation does not guarantee that the authority is the one governing the case.

Significance and Contributions
The paper makes three primary contributions:

Identification of a Validity Problem: It formalizes the observation that answer-only scoring treats naturally occurring gold-authority misses as complete successes. This is not an adversarial stress-test failure but a systematic issue within ordinary benchmark regimes.
Joint Scoring Protocol: The author proposes a minimal, low-cost extension to existing statute-grounded benchmarks. Instead of a single accuracy metric, evaluation should report:
- Answer accuracy.
- Authority Hit@k (grounding rate).
- The joint distribution ( $A^+G^+, A^+G^-, A^-G^+, A^-G^-$ ).
Methodological Shift: The study argues that "fluent legal form" (correct reasoning text with citations) cannot be treated as evidence that the asserted authority supports the conclusion. Because statutory authority is externally verifiable, benchmarks must audit the link between the answer and the cited law, not just the answer itself.

Limitations and Scope
The author notes that their claims are behavioral, not mechanistic; the study measures the output mismatch, not the internal reasoning process. The findings are most directly applicable to codified, citation-normed tasks (like the Taiwan Bar Exam). The magnitude of the mismatch might differ in open-ended legal tasks or common-law settings where specific article citations are less rigid. Additionally, the study includes a preliminary observation of similar spontaneous authority marking in Mainland Chinese civil law, motivating future cross-jurisdictional audits, but does not yet provide a full joint analysis for that dataset.

Technical Summary: Knowing the Form, Not the Function

More like this