Research·arXiv cs.CL·16h ago

Two-agent system enforces per-claim verification in legal AI outputs

Researchers introduce GANDR, a two-agent verification system that audits language model outputs claim-by-claim rather than scoring answers holistically. In legal and other high-stakes domains, this addresses a critical failure mode where models generate plausible-sounding conclusions supported by fabricated or mismatched citations. The Drafter produces structured reasoning while a Critic independently validates each assertion against source material, producing per-claim audit traces. This work signals growing recognition that end-to-end answer quality metrics mask citation hallucination, a persistent problem undermining LLM deployment in regulated fields where verifiability is non-negotiable.

Modelwire context

Explainer

GANDR's key insight is architectural rather than just methodological: it treats verification as a separate agent task with its own audit trail, rather than folding it into a single scoring function. This separation makes the hallucination visible and traceable in a way end-to-end metrics cannot.

This work sits squarely within a pattern we've tracked across multiple papers this week. Like the symbolic-reasoning framework from earlier today that decomposes visual problems into explicit logic steps, and TRACE's approach to synthesizing verifiable rewards in causal reasoning, GANDR treats verifiability as a first-class architectural concern rather than a post-hoc patch. The common thread: when stakes are high (legal reasoning, diagnostic AI, formal proof), decomposing the problem into independently auditable stages beats unified end-to-end models. The difference here is that GANDR targets a specific failure mode (citation hallucination) that holistic metrics systematically miss, making it directly applicable to regulated deployment scenarios where citation provenance is non-negotiable.

If GANDR's per-claim audit traces are adopted in production legal AI systems within the next 18 months (watch for announcements from Thomson Reuters, LexisNexis, or similar), that signals the field has moved past treating hallucination as an inherent LLM property toward treating it as a solvable architectural problem. If adoption stalls and vendors continue shipping holistic confidence scores instead, that tells you the cost of per-claim verification exceeds the regulatory pressure to eliminate it.

Coverage we drew on

From Symbolic Perception to Logical Deduction: A Framework for Guiding Language Models in Geometric Reasoning · arXiv cs.CL

This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.

MentionsGANDR

Read full story at arXiv cs.CL →(arxiv.org)

Modelwire Editorial

This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.

Our mission How we write

Modelwire summarizes, we don’t republish. arXiv cs.CL originally reported this story as “GANDR: Claim Auditing for Verifiable Legal Answer Generation”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.