MUMBAI, India, May 29 -- Intellectual Property India has published a patent application (202611051146 A) filed by Noida Institute Of Engineering And Technology, Greater Noida, Uttar Pradesh, on April 22, for 'a hardware-accelerated system and method for value-aware numerical representation in transformer-based language models.'
Inventor(s) include Manish Chaudhary; and Priya Dahiya.
The application for the patent was published on May 29, under issue no. 22/2026.
According to the abstract released by the Intellectual Property India: "The present invention relates to a hardware-accelerated system (100) and method for value-aware numerical representation in transformer-based language models. The system includes a processor unit (101), memory subsystem (102), and a tokenizer module (103) for parsing numerical expressions. A numerical value extraction unit (104) derives underlying magnitudes, while a value-conditioned embedding generator (105), comprising a multilayer perceptron (106) and recurrent neural network (107), produces value-aware embeddings. A transformer engine (108) with multi-head self-attention layers (109) processes sequences augmented with special numerical tokens (110) encoding magnitude information. A projection module (111) aligns hidden states with value-based embeddings, and a composite loss unit (112) optimizes reconstruction accuracy. The system enhances arithmetic reasoning and numerical robustness in language model applications."
Disclaimer: Curated by HT Syndication.
