Large language model dynamic guiding method based on condition generation model and related device

By dynamically generating guiding vectors through a conditional generative model, the fragility and instability of guiding vectors in large language models are resolved, adaptive guiding control is achieved, and the stability and automation level of the model are improved.

CN122133664APending Publication Date: 2026-06-02XI AN JIAOTONG UNIV
0 Cites 0 Cited by

Patent Information

Application Number
CN202610201046.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-02-11
Publication Date
2026-06-02

AI Technical Summary

Technical Problem

Existing guided vector methods for large language models suffer from fragility, instability, lack of adaptability, and reliance on manual selection, resulting in unstable model behavior and poor generalization ability.

Method used

A dynamic guidance method based on a conditional generative model is adopted. By obtaining the activation vector of the user prompt, it is projected onto the pre-extracted task semantic base, and the target task is identified by geometric matching. The guidance vector is dynamically generated by the conditional generative model and injected into the residual stream to achieve adaptive guidance control.

Benefits of technology

It improves the stability and automation of large language model guidance, reduces storage overhead, and enhances the scalability and robustness of the system.

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

Abstract

The invention belongs to the technical field of electronic information, and discloses a large language model dynamic guiding method based on a condition generation model and a related device. The large language model dynamic guiding method comprises the steps that a user prompt is obtained, and an activation vector of the user prompt is obtained through a designated intervention layer of a large language model; projecting the activation vector to each pre-extracted task semantic base, and identifying a target task to which the user prompt belongs through geometric matching; and taking the target task as an input condition, dynamically generating a guide vector through a pre-trained condition generation model, and injecting the guide vector into a residual flow of a large language model for dynamic guide. According to the technical scheme disclosed by the invention, the stability and the automation level of large language model guidance can be greatly improved, and meanwhile, the generalization ability and the robustness can also be improved.
Need to check novelty before this filing date? Find Prior Art