MUMBAI, India, June 30 -- Intellectual Property India has published a patent application (202641075838 A) filed by Vellore Institute Of Technology on June 19, 2026, for System And Method For Training-Free Semantic Drift Correction Using Anchor Refresh Decoding.
Inventors include Priya G; and Aditya Sudhakar.
The application for the patent was published on June 26, 2026, under issue no. 26/2026.
Abstract: ABSTRACT The present invention relates to a system and method for maintaining semantic coherence during long-form text generation by autoregressive language models. The invention comprises an Input Prompt (101), a Topic Extraction module (102), a Context Initialization module (103), a Sentence Generation module (104), a Drift Detection module (105), an Anchor Injection module (106), and a Context Update module (107). A topic phrase is extracted from an input prompt and encoded into a semantic embedding that serves as a topic anchor throughout a generation session. Generated sentences are continuously evaluated by computing a similarity score between sentence embeddings and the topic anchor. Upon detection of semantic drift, a compressed topic reminder is injected into the active context and decoding parameters are temporarily adjusted by narrowing a nucleus sampling range to promote topic-relevant token generation. The invention performs real-time drift correction without retraining, model modification, external retrieval systems, or GPU acceleration, thereby enabling coherent and resource-efficient long-form text generation on CPU-based computing platforms.
Disclaimer: Curated by HT Syndication.
