Intellectual Property India Publishes Patent Application for A Method And System For Context-Aware Hallucination Detection In Large Language Models Using Contrastive Semantic Anchoring

MUMBAI, India, July 7 -- Intellectual Property India has published a patent application (202621058065 A) filed by Dr. Prashant G Ahire; Dr. Nidhi Dandotiya; Dr. Rahul A Patil; Dr. Deepak Gupta; C P Bhargava; Dr. Vaibhav Nivrutti Patil; Tukaram Dethe; Dr. Pushpa Gopal Ambhore; Dr. Pradeep Yadav; and Dr. Abhinandan Singh Dandotiya on May 07, 2026, for A Method And System For Context-Aware Hallucination Detection In Large Language Models Using Contrastive Semantic Anchoring.

Inventors include Dr. Prashant G Ahire; Dr. Rahul A Patil; Dr. Deepak Gupta; C P Bhargava; Dr. Vaibhav Nivrutti Patil; Tukaram Dethe; Dr. Pushpa Gopal Ambhore; Dr. Pradeep Yadav; Dr. Abhinandan Singh Dandotiya; and Dr. Nidhi Dandotiya.

The application for the patent was published on July 03, 2026, under issue no. 27/2026.

Abstract: The present invention provides a method, system, and non-transitory computer-readable medium for context-aware hallucination detection in large language model (LLM) outputs using Contrastive Semantic Anchoring (CSA). The invention segments LLM- generated responses into atomic claim units (ACUs) via a Claim Segmentation Pipeline, constructs a semantic anchor vector space from verified contextual passages using a contrastively pre-trained sentence encoder, and computes Contrastive Semantic Divergence (CSD) scores representing semantic dissimilarity between each ACU and the anchor centroid. A Hallucination Classification Module applies a learned threshold to produce binary hallucination labels and confidence scores per ACU, yielding a structured hallucination report with aggregate risk assessment. The system supports real-time streaming operation, domain-adaptive fine-tuning via LoRA adapters, multi-modal anchor construction, and integration with LLM decoding as a constrained generation mechanism. The invention demonstrates superior hallucination detection precision and recall compared to prior art methods on benchmark datasets and is industrially applicable across medical, legal, financial, and e-governance domains. FIG 01.

Disclaimer: Curated by HT Syndication.