A large language model misleading information detection and correction system

Abstract

This invention provides a system for detecting and correcting misleading information in a large language model, comprising: integrating a state machine processing engine into a deterministic finite state automaton-based scheduling process, where state transitions are driven by the attributes of the input question and knowledge state signals; designing a dynamic knowledge update unit; continuously crawling knowledge from external authoritative sources using a lightweight knowledge crawler; calculating knowledge freshness scores to drive the process; and maintaining the timeliness of the large language model knowledge base through an incremental knowledge injection mechanism; sequentially scheduling question detection, question correction, and question answering stages based on the state machine processing engine and the dynamic knowledge update unit, iteratively updating until a final accurate response is generated. This invention combines theoretical rigor with real-time knowledge, mathematically ensuring the correctness of model behavior and dynamically integrating the latest knowledge, thereby providing continuous and reliable services in complex and ever-changing real-world scenarios.

Images