Intellectual Property India Publishes Patent Application for 'System And Method Of Compiler-Orchestrated Validation And Optimization Framework For Large And Small Language Models' Filed by Persistent Systems

MUMBAI, India, Jan. 23 -- Intellectual Property India has published a patent application (202521095661 A) filed by Persistent Systems, Pune, Maharashtra, on Oct. 6, 2025, for 'system and method of compiler-orchestrated validation and optimization framework for large and small language models.'

Inventor(s) include Mr. Nitish Shrivastava; and Mr. Pradeepkumar Sharma.

The application for the patent was published on Jan. 23, under issue no. 04/2026.

According to the abstract released by the Intellectual Property India: "The present invention provides a computer-implemented system and method for orchestrating task execution using a large language model (LLM) and a small language model (LLM) without requiring retraining. The system comprises functional modules including a Query Clarification Module to parse and refine user queries through a limited number of clarifying questions, a Task Decomposition Module to partition refined queries into atomic subtasks, and a Large LLM Specification Module to generate per-subtask specifications comprising YAML-based rules, validation workflows, test cases, executable validation scripts, and detailed prompts. A Small LLM Execution Module executes the prompts, runs validations, and produces errors and warnings. An Orchestrator Module iteratively refines specifications in response to diagnostics until convergence. An Output Generator Module produces a structured directory containing validated prompts, rules, workflows, scripts, and test cases. This runtime orchestration approach enables the small LLM to achieve near-large-model quality, offering a robust, portable, and auditable alternative to conventional model distillation or retraining."

Disclaimer: Curated by HT Syndication.