MUMBAI, India, Sept. 14 -- Intellectual Property India has published a patent application (202621036764 A) filed by Persistent Systems on March 26, 2026, for System And Method For Response Gap Reconciliation Arising From Input Obfuscation.
Inventor includes Mr. Nitish Shrivastava.
The application for the patent was published on September 11, 2026, under issue no. 37/2026.
Abstract: A system [300] and method for detecting and reconciling semantic gaps arising in large language model responses generated from obfuscated inputs. The system [300] implements sliding window coherence analysis, combining perplexity-based and entailment-based dual metrics to identify discontinuities where semantic content should logically appear but does not. A gap classifier [330] employs five hierarchical categories to characterize identified discontinuities: referential inconsistency [331], factual misalignment [332], contextual incoherence [333], hallucinated elaboration [334], and stylistic boundary artifacts [335]. Local small language model reconciliation generators [340] apply category-specific prompt templates tailored to repair mechanisms for each discontinuity type. The system [300] executes multi- pass coherence verification with configurable delivery thresholds to ensure restored content maintains semantic integrity. Detection of referential inconsistencies utilizes coreference resolution and pronoun property extraction. Factual misalignment detection employs information extraction models to identify quantitative assertions and cross-reference them against obfuscated value constraints.
Disclaimer: Curated by HT Syndication.
