MUMBAI, India, Aug. 31 -- Intellectual Property India has published a patent application (202611072433 A) filed by Noida Institute Of Engineering And Technology Niet on June 11, 2026, for Hardware-Accelerated Knowledge-Graph-Grounded Reward Computation System For Compositional Multi-Hop Reasoning In Language Models.
Inventors include Pradeep Kumar; and Maneesh Kumar.
The application for the patent was published on August 28, 2026, under issue no. 35/2026.
Abstract: A hardware-accelerated knowledge-graph-grounded reward computation system (100) for enabling compositional multi-hop reasoning during reinforcement-learning-based post-training of language models. The system (100) comprises a knowledge graph storage subsystem (110) with a triple store memory (112) protected by an IOMMU controller (114) and a write-lock latch register (116) that physically latches the memory to a read-only state, a path traversal engine (120) with a hop-counter finite state machine (122) and an entity matcher circuit (124) and a coverage accumulator (126) for computing path alignment between a model-generated reasoning trace and ground-truth knowledge graph paths, a reward computation unit (130) for combining binary correctness reward with path alignment reward, a bus interconnect firewall (170) with latched master-identifier filter registers for isolating the reward computation datapath, and a commit controller (180) with a monotonic counter for creating an immutable audit trail of reward computations.
Disclaimer: Curated by HT Syndication.
