MUMBAI, India, Sept. 22 -- Intellectual Property India has published a patent application (202641108531 A) filed by Vellore Institute Of Technology on September 10, 2026, for System And Method For Closed-Loop Inference Control And Dynamic Threshold Recalibration Of Transformer-Based Language Models.
Inventors include D. Sumathi; Deepak C; Dhanush T; Dhanush M; and Kishore E.
The application for the patent was published on September 18, 2026, under issue no. 38/2026.
Abstract: A system for AI-based adaptive multi-hazard safety assessment and prioritization in smart-home environments is provided. The system 100 includes user device 102, network 104, and inference-control server 106 to process an input prompt and determine, at a plurality of token positions, a prior-dominance score based on a ratio between entropy of a prior distribution and entropy of a context-conditioned distribution. When the score exceeds an operative domain-specific threshold, the server 106 identifies attention heads and feed-forward regions and injects a domain- and severity-based correction vector into a residual stream. A directed acyclic graph representing a multi-step reasoning output is constructed to determine a causal-fidelity score. Upon failed verification, correction is repeated at the same residual-stream location for a bounded number of attempts with escalating suppression gain and attempt-conditioned vector re-selection. Control-decision records are cryptographically hash-linked and Merkle-committed, enabling cross-model-scale analysis, monotonicity-violation determination, bounded threshold recalibration, and checkpoint rollback. FIG. 1
Disclaimer: Curated by HT Syndication.
