LLMs Churning Out Fake Legal Citations: Finally a Fix?

Legal professionals, brace yourselves. Large language models (LLMs) are proving unreliable legal citations. They're not just occasionally off. they're often downright fabricating. With references to repealed statutes and jurisdictional mix-ups, it's a mess. But there's a glimmer of hope with a new metric called citation grounding (CG), which might just clean up this chaotic legal citation landscape.

Citation Grounding: A Breakthrough?

CG takes a deep dive into the LLM-generated citations and verifies them against a massive database of Ukrainian court decisions. We're talking about a citation graph with 100.8 million court decisions and 502 million edges. That's no small feat. CG splits into three parts: checking if a cited provision actually exists, assessing its contextual relevance, and verifying if it was valid at the right time. This allows for identifying exactly where the hallucinations are happening.

The testing ground was 100 Ukrainian legal queries across five systems, including four commercial LLMs from AWS Bedrock and a RAG-augmented production system. Results ranged from CG scores of 0.791 to 0.873. But here's the kicker: 13-21% of those citations were hallucinated. It's not perfect, but it's a start.

Automating the Fix

To tackle this without endless human oversight, there's a new method in town: Citation Grounding DPO (CG-DPO). This technique creates preference pairs by deliberately corrupting verified citations, offering a training set for models. On a dataset of 2,244 court decisions, a Qwen2.5-7B-Instruct model with LoRA hit a 98.5% accuracy in telling correct from corrupted. That's a reward margin of +14.9, with a tiny standard deviation under 0.3 percentage points across three trials.

Why does this matter? Because legal systems rely on accuracy. Hallucinated citations can lead to misguided decisions, and that's a big problem. But is CG-DPO the golden ticket? Or just another technical band-aid?

Show Me the Results

The resources, including the citation graph and evaluation framework, are open-access. That's a promising step. But what's really needed is transparency and widespread adoption. Will legal professionals trust these AI fixes? Will these methods hold up across different jurisdictions and legal systems?

The reality is, without tangible improvements in citation accuracy, LLMs will remain unreliable for serious legal work. It's a promising start, but show me the product's impact on real-world legal accuracy. Until then, it's just vaporware.