MUMBAI, India, July 30 -- Intellectual Property India has published a patent application (202641087980 A) filed by Christ University on July 18, 2026, for System And Method For Cascaded Large Language Model Inference With Failure Summarization And Token Budgeting.
Inventors include Dr. Nipuna Ashok; Dr. Jossy George; and Dr. Monisha Singh.
The application for the patent was published on July 24, 2026, under issue no. 30/2026.
Abstract: The invention relates to machine learning inference for code generation, particularly a cascaded large language model architecture that improves compute efficiency and correction quality over single-stage inference. An original code-generation task is first processed by a lightweight 1.5B-parameter coding-oriented instruction model under a limited token budget, after which an execution judge deterministically validates the generated code in an isolated sandbox using hidden tests, tracing, a five- second hard limit, and loop and contamination controls. On failure, sanitized traceback or exception state information is compressed into a semantic diagnostic summary that preserves corrective error semantics while removing redundant code, warnings, traceback text, and reasoning trajectory. The original task and summary are then supplied, without failed code output, to a 7B-parameter instruction model for repair-oriented generation under a separate token budget. The invention is particularly useful for Python programming tasks judged through hidden execution tests.
Disclaimer: Curated by HT Syndication.
