Intellectual Property India Publishes Patent Application for A Hierarchical Multi-Agent System With Rule-Based Scoring For Real-Time Hallucination Detection And Iterative Correction In Large Language Model Deployments

MUMBAI, India, July 7 -- Intellectual Property India has published a patent application (202641076688 A) filed by Dr V. Sidda Reddy; Dr S. A. Poojitha; Ms. Ishrat Ruquia; Ms. Gorak Tanusha; Ms. B. Sandhya; and Ms. M. Tulasi on June 20, 2026, for A Hierarchical Multi-Agent System With Rule-Based Scoring For Real-Time Hallucination Detection And Iterative Correction In Large Language Model Deployments.

Inventors include Dr V. Sidda Reddy; Dr S. A. Poojitha; Ms. Ishrat Ruquia; Ms. Gorak Tanusha; Ms. B. Sandhya; and Ms. M. Tulasi.

The application for the patent was published on July 03, 2026, under issue no. 27/2026.

Abstract: The present invention discloses a hierarchical multi-agent framework for real-time detection and iterative correction of hallucinations generated by Large Language Models (LLMs) in production-scale deployments. The system comprises a Parent Orchestrator Agent that decomposes complex tasks into atomic sub-tasks, each managed by a dedicated Consultant-Evaluator agent pair. The Consultant Agent, implemented using an open-source LLM such as LLaMA-3-8B, generates initial task outputs, while the Evaluator Agent, implemented using Mistral-7B, invokes a deterministic rule-based scoring function based on the Ratcliff/Obershelp Gestalt pattern matching algorithm to compute a similarity score (Dro = 2Km / (|S1| + |S2|)) between generated output chunks and source reference segments. Chunks with similarity scores below an 80% threshold are flagged as hallucinations and returned to the Consultant Agent for correction within a finite-iteration control loop (maximum 2 iterations). A Performance Database logs hallucination rates, similarity scores with 95% confidence intervals, and agent performance metrics for continuous monitoring and anomaly detection. Applied to 3,680 dialogue chunks from 308 real-world call center transcripts, the system achieves an 85.5% reduction in hallucination rate for LLaMA-3-8B and 67.7% for Mistral-7B, demonstrating its efficacy for scalable, domain-agnostic LLM deployment in applications requiring high factual accuracy.

Disclaimer: Curated by HT Syndication.