Learning method and device during large language model testing, equipment and medium

By extracting the lexical-level evidence quality and cognitive uncertainty of the large language model generation steps, and combining low-rank adaptation technology to construct a mask for the model update window, the problem of uncertainty estimation and unstable update of large language models in the distribution offset scenario is solved, achieving more efficient and accurate model adaptation and robustness.

CN122287755AActive Publication Date: 2026-06-26CHONGQING CHANGAN AUTOMOBILE CO LTD
9 Cites 0 Cited by

Patent Information

Application Number
CN202610776450.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-06-01
Publication Date
2026-06-26
Estimated Expiration
2046-06-01

AI Technical Summary

Technical Problem

Existing large language models suffer from inaccurate uncertainty estimation, unstable update process, simplistic loss function design, and an imbalance between computational efficiency and adaptation effect in distributed offset scenarios, resulting in insufficient reliability and robustness of the models in real-world complex environments.

Method used

By extracting the lexical-level evidence quality from the generation step, calculating and smoothing cognitive uncertainty, constructing a mask for the model update window, combining evidence quality-weighted loss, and using low-rank adaptation technology to update parameters, stable regions are selected for updating.

Benefits of technology

It improves the adaptability and robustness of large language models in distributed offset scenarios, reduces computational overhead, enhances the model's anti-interference ability, prevents adaptive drift and error accumulation, and improves the efficiency and accuracy of learning during testing.

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

Abstract

The invention discloses a learning method, device and equipment during large language model testing and a medium, and the method comprises the steps: inputting test data into a large language model with low-rank adaptation parameters, and obtaining a prediction score of a lexical element in a vocabulary in each generation step; based on the prediction score, lexical-level evidence quality of the generation step is extracted, and lexical-level cognitive uncertainty of the generation step is calculated; smoothing the lexical element level cognitive uncertainty, determining a stable region of the generation step according to the smoothing uncertainty, and constructing a mask of a model update window; and calculating mask entropy loss based on the mask, and updating the low-rank adaptation parameter through back propagation. According to the invention, the problem of self-adaptive drift in model updating in the prior art can be solved.
Need to check novelty before this filing date? Find Prior Art