MUMBAI, India, Sept. 22 -- Intellectual Property India has published a patent application (202641108529 A) filed by Vellore Institute Of Technology on September 10, 2026, for System And Method For Low-Latency Semantic Communication Between Heterogeneous Neural Networks.
Inventors include Kalaiselvan N; and Shreejeet Chowdhury.
The application for the patent was published on September 18, 2026, under issue no. 38/2026.
Abstract: Embodiments herein provide a system (100) for low-latency semantic communication between heterogeneous neural networks. The system (100) includes a first processor (106) executing a first neural network (104), a second processor (130) executing a second neural network (128), a memory (108, 132), an inter-agent communication system (112), and a communication interface (110, 124). The first processor (106) captures a final hidden state generated before language decoding as a latent vector and stores the latent vector with a deterministic TTL and a cryptographic pointer (206). The communication interface (110, 124) transmits the cryptographic pointer (206) instead of serialized natural-language information. The second processor (130) retrieves the latent vector, transforms the latent vector using a pre-computed SVD-based latent translation matrix when required, and directly injects the latent vector or transformed latent vector into a KV cache (134) without lexical tokenization or embedding lookup. FIG. 1
Disclaimer: Curated by HT Syndication.
