MUMBAI, India, Sept. 28 -- Intellectual Property India has published a patent application (202641113544 A) filed by Vellore Institute Of Technology on September 22, 2026, for System And Method For Perturbation-Based Detection, Quantification, And Control Of Source-Preference Sycophancy In Retrieval- Augmented Generation.
Inventors include Vani Rajasekar; Nikhil Agarwal; and Sonone Aryan Swapnil.
The application for the patent was published on September 25, 2026, under issue no. 39/2026.
Abstract: The present disclosure proposes a system (100) and method for detecting, quantifying, and controlling source-preference sycophancy in a target large language model (102) operating within a retrieval-augmented generation pipeline (104). The system (100) includes a computing device (106), a processor (108), a memory (110), and plurality of modules (112) configured for conflicting-document construction, controlled authority-credential and temporal-conviction perturbation, factorial source-label and content-truth variation, output-alignment classification, reasoning-output divergence detection, and sensitivity aggregation. A sensitivity aggregation module (126) generates a composite sycophancy-sensitivity score from authority-credential sensitivity, temporal-framing sensitivity, and reasoning-output divergence information. A gating module (128) compares derived sensitivity information with a configurable threshold. A runtime interface (130) supports response control by preventing release of a generated output or causing release of a replacement output before delivery to an end user.
Disclaimer: Curated by HT Syndication.
