What is it about?
When an AI system queries a knowledge graph and gets the right facts but assembles them in the wrong order: "Activision acquired Microsoft" instead of the reverse, standard detection methods fail completely. They look at words, not structure. ReGA takes a different approach: it converts both the trusted source and the AI's answer into small graphs, then checks whether the relationships point in the right directions. A fast CPU stage handles most cases in 14 milliseconds. A neural graph matcher resolves the hard cases in 3 milliseconds. Together they are 88 times faster than asking a large language model to verify the same thing, with no ongoing cost after a one-time 7-second training step.
Featured Image
Photo by GuerrillaBuzz on Unsplash
Why is it important?
Production AI systems that answer questions over structured data: search engines, enterprise knowledge bases, legal and medical tools, cannot afford to wait one second per verification. ReGA makes structural fact-checking fast enough to run inside the generation loop, catching and correcting errors before the user ever sees them. The legal domain result (AUC 0.984) is particularly relevant: in high-stakes settings where getting a relationship direction wrong has real consequences, ReGA matches a 70B language judge at 200 times lower latency.
Perspectives
We built this because standard hallucination metrics completely miss relational inversions. "Activision acquired Microsoft" scores fine on every text-based measure. Once you treat it as a graph problem instead of a language problem, the error is trivial to detect. The surprise was that random node embeddings outperform semantic ones, which tells you the problem is purely structural. We have been overcomplicating a class of verification problems that are, at their core, just graph isomorphism.
Valentin NOËL
Devoteam
Read the Original
This page is a summary of: ReGA: Zero-Overhead Graph Alignment for Structural Hallucination Detection Without Generation, May 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3774905.3794657.
You can read the full text:
Contributors
The following have contributed to this page
