Large model detection method and device, storage medium and electronic equipment

By injecting the sectional program record logs in the big model business process and detecting its execution process, the problem of big model illusion is solved and the output credibility in key scenarios is improved.

CN120429137APending Publication Date: 2025-08-05ALIPAY (HANGZHOU) INFORMATION TECH CO LTD
0 Cites 0 Cited by

Patent Information

Application Number
CN202510414013.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-04-02
Publication Date
2025-08-05

AI Technical Summary

Technical Problem

When executing business, large models may have hallucinations due to errors in data or training process errors, resulting in their generation of descriptions that are logically reasonable but actually do not conform to known common sense, especially in scenarios such as law, medical care, and finance that may have a significant impact on user interests or security.

Method used

The first sectional program is pre-injected into the external program required to be called when the big model executes the business, record the execution logs, and obtain and analyze these logs after the business is completed through the second sectional program to determine the execution process of the big model, thereby detecting whether it has hallucinations.

Benefits of technology

By recording and analyzing the call logs of external programs, we can accurately determine whether the big model has hallucinations when executing business, which improves the credibility of the output results and ensures security and accuracy in key scenarios.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The embodiment of the invention discloses a detection method of a large model, which comprises the following steps of: injecting a first section program into an external program required for executing a service in advance to record an execution log when the external program is called by the large model, and after the large model executes the service, executing the external program through the execution log recorded by the first section program. According to the method, the calling execution process of the large model on each external program when the large model executes the business is determined, and whether illusion occurs in the large model in the business execution process is detected according to the execution process, so that the credibility of an output result of the large model is judged.
Need to check novelty before this filing date? Find Prior Art