MUMBAI, India, June 26 -- Intellectual Property India has published a patent application (202641072843 A) filed by Vellore Institute Of Technology on June 12, 2026, for Hallucination Detection In Large Language Model Outputs.
Inventors include Dr. Priyadharsi Ni M; Unta Chaitanya; and Milan Rose Joshy.
The application for the patent was published on June 19, 2026, under issue no. 25/2026.
Abstract: ABSTRACT HALLUCINATION DETECTION IN LARGE LANGUAGE MODEL OUTPUTS A method for detecting hallucinations in outputs generated by a large language model (200) comprises: receiving a text output from the large language model (200); extracting atomic claims from the text output, each classified into a claim type; constructing a directed dependency graph representing logical dependencies among the atomic claims; computing a verification priority score for each atomic claim based on domain risk weight, structural centrality score, and uncertainty score; retrieving evidence from heterogeneous knowledge sources for each atomic claim; generating a fused belief distribution by combining evidence using Dempster-Shafer evidential reasoning; producing a calibrated confidence score and uncertainty interval for each atomic claim; classifying each atomic claim into confidence tiers based on the calibrated confidence score; and propagating confidence penalties from atomic claims in lower confidence tiers to dependent atomic claims in the directed dependency graph. (Fig. 1)
Disclaimer: Curated by HT Syndication.
