MUMBAI, India, July 30 -- Intellectual Property India has published a patent application (202641088136 A) filed by Vellore Institute Of Technology on July 20, 2026, for System And Method For Adaptive Reasoning Optimization In Large Language Model Inference.
Inventors include Sanjiban Sekhar Roy; and Malpure Darshan Vasudev.
The application for the patent was published on July 24, 2026, under issue no. 30/2026.
Abstract: ABSTRACT The present invention relates to a computer-implemented system and method for adaptive reasoning optimization in large language model (LLM) inference. An input assessment question (101) is received and classified by a question category classifier (102) into quantitative reasoning, logical reasoning, or situational judgment categories. Based on the classification, a strategy router selectively applies a quantitative chain-of-draft mode (103), a logical chain-of-thought mode (104), or a situational direct-answer mode (105). A token budget safety gate prevents deployment of verbosity-constrained reasoning modes when a predefined accuracy threshold would be violated, thereby preserving correctness for accuracy-critical categories. The system further comprises a dynamic telemetry-aware constraint calculator configured to adjust reasoning verbosity based on runtime conditions. An LLM API interface generates a corresponding response (106). The invention operates entirely through prompt-level modifications at inference time without model fine-tuning, weight modification, training data changes, or specialized hardware, thereby reducing token consumption, inference cost, and latency while maintaining or improving prediction accuracy.
Disclaimer: Curated by HT Syndication.
