MUMBAI, India, Sept. 14 -- Intellectual Property India has published a patent application (202611086561 A) filed by Noida Institute Of Engineering And Technology Niet on July 15, 2026, for Tokenization-Aware Safety Enforcement System And Method For Large Language Models.
Inventors include Dr. Hitesh Singh; and Anamika Srivastav.
The application for the patent was published on September 11, 2026, under issue no. 37/2026.
Abstract: A tokenization-aware safety enforcement apparatus (100) for large language models implemented on computing devices comprising processors, memory, and storage. The apparatus comprises a multi-strategy canonicalization engine (120) that reconstructs canonical token representations from adversarially perturbed inputs using bigram frequency analysis, a token fragmentation detector (130) computing a composite fragmentation score using a dynamic safety vocabulary, a dual-path safety evaluator (140) simultaneously processing canonical and raw representations to detect safety divergence, an adaptive training data augmentation engine (150) generating fragmented safety training examples through six perturbation families with curriculum-based injection, and a refusal calibration module (160) employing trained linear probes coupled to neural network layer activations for robust refusal decisions. The apparatus produces the technical effect of preventing adversarial exploitation of subword tokenization boundaries in computing infrastructure, reducing attack success rate while maintaining over-refusal below ten percent. [Figure 1]
Disclaimer: Curated by HT Syndication.
