Large model multi-language security alignment system and method based on joint rejection direction

By constructing multilingual secure parallel samples and value stream masking technology, and extracting joint rejection directions shared across languages, the problem of insecure response of low-resource language models to malicious requests is solved, and efficient security rejection capability and cross-language secure transfer are achieved.

CN122451091APending Publication Date: 2026-07-24HARBIN INST OF TECH
0 Cites 0 Cited by

Patent Information

Application Number
CN202610823295.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-06-09
Publication Date
2026-07-24

AI Technical Summary

Technical Problem

Existing technologies struggle to effectively enhance the secure rejection capability of large language models under low-resource language conditions. Furthermore, traditional methods lack cross-language secure transfer mechanisms, struggle to provide interpretable secure representations and efficient parameter alignment, leading to insecure responses from low-resource language models when faced with malicious requests.

Method used

By constructing multilingual security parallel samples, constructing positive and negative security states using value flow masks, extracting joint rejection directions shared across languages, and aligning through layer-by-layer projection trajectories and lightweight parameter learning, the rejection trajectories of low-resource languages ​​are made to approximate the rejection trajectories of high-resource anchored languages, thereby achieving internal characterization and alignment of security behaviors.

Benefits of technology

This study improves the safe rejection capability of low-resource languages ​​under limited sample conditions, maintains the original general capabilities of the model, achieves robust deep cross-language safe transfer, and ensures that low-resource languages ​​generate compliant rejection responses when faced with malicious intent.

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

Abstract

The invention relates to a large model multi-language security alignment system and method based on a joint rejection direction. The invention relates to the technical field of large model security alignment, and discloses a multi-language security parallel sample construction method, which constructs a positive security state and a negative security state by using a value flow mask on a frozen large language model, and further extracts a cross-language shared joint rejection direction from the internal representation difference of the positive security state and the negative security state. And taking a layer-by-layer projection track in the direction as an internal description of a safety behavior, and enabling a track of a low-resource language to approach a high-resource anchoring language through lightweight parameter learning, so that the safety rejection capability of the low-resource language is improved under the condition of few samples, and the original universal capability of the model is kept as much as possible.
Need to check novelty before this filing date? Find Prior Art