MUMBAI, India, Aug. 12 -- Intellectual Property India has published a patent application (202621036763 A) filed by Persistent Systems on March 26, 2026, for A System And Method For Semantic Equivalence Preservation During Large Language Model Inference.
Inventor includes Mr. Nitish Shrivastava.
The application for the patent was published on August 07, 2026, under issue no. 32/2026.
Abstract: The present invention relates to a Semantic Equivalence Preservation Engine (SEPE) system (200) and method generating context-aware, privacy-preserving placeholder replacements during large language model inference while maintaining semantic consistency and relational coherence. The SEPE system architecture (200) comprises a category classifier module (210) configured to classify detected entities into predefined categories, a candidate generation pipeline (220) including category-specific generators (221a–221h) configured to generate semantically appropriate replacement candidates, and an embedding-based similarity selector module (230) evaluating candidates using contextual embeddings and similarity thresholds. A reverse sensitivity checker module (240) validates selected replacements against sensitive entity databases and pattern detection mechanisms to prevent re-identification. A hierarchical fallback module (250) applies progressive degradation strategies when suitable candidates are unavailable, and a relational consistency enforcer module (260) preserves mathematical, temporal, and co-reference relationships between replaced entities. The system operates using local small language model (SLM) (205) for on-device processing without external data transmission.
Disclaimer: Curated by HT Syndication.
