MUMBAI, India, Sept. 22 -- Intellectual Property India has published a patent application (202641110074 A) filed by Vellore Institute Of Technology on September 14, 2026, for Black-Box Causal Prompt Compression System With Rate-Distortion Encoding And Cross-Model Consensus Validation.
Inventors include Jagadeesan S; Dhanush M; Deepak C; Vishva S A; and Dhanush T.
The application for the patent was published on September 18, 2026, under issue no. 38/2026.
Abstract: ABSTRACT BLACK-BOX CAUSAL PROMPT COMPRESSION SYSTEM WITH RATE-DISTORTION ENCODING AND CROSS-MODEL CONSENSUS VALIDATION A computer-implemented system (100) for compressing an input prompt for a large language model accessed through an external inference application programming interface (114) is provided. The system (100) comprises a Causal Query Scheduler (102) configured to allocate a finite, pre-declared query budget across a plurality of token-masking interventions by iteratively eliminating candidate tokens according to a successive-halving best-arm-identification schedule based on an upper-confidence causal-effect estimate. The system (100) further comprises a Potential-Outcome Estimator (104) configured to compute, for each surviving candidate token, a doubly-robust estimate of a conditional average treatment effect from paired input/output observations obtained through the external inference application programming interface (114), without access to attention weights, gradients, or parameters of the large language model. The system (100) further comprises a Black- Box Causal Influence Graph constructor (106), a Rate-Distortion Prompt Encoder (108), a Multi-Agent Compression Negotiation Engine (110) with a Coordination Policy Network, and a Cross-Model Consensus Validator (112) comprising a plurality of heterogeneous proxy models and a reliability-weighted voting module configured to accept or reject the candidate compressed prompt based on a consensus score computed from historically-tracked, production-outcome-updated per-proxy reliability weights.
Disclaimer: Curated by HT Syndication.
