Reinforcement learning role large model personality consistency alignment optimization method and related products

By using reinforcement learning and closed-loop control, the personality bias of large pre-trained language models is monitored and corrected in real time, which solves the problem of personality inconsistency in long dialogues and improves the consistency and stability of the model's performance in multi-turn dialogues.

CN121745210BActive Publication Date: 2026-07-03LIANGSHENG DIGITAL CREATIVE DESIGN (HANGZHOU) CO LTD
View PDF 1 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
LIANGSHENG DIGITAL CREATIVE DESIGN (HANGZHOU) CO LTD
Filing Date
2026-02-28
Publication Date
2026-07-03

AI Technical Summary

Technical Problem

In existing technologies, large-scale pre-trained language models are prone to personality drift and inconsistency in long-term, multi-turn dialogues. The lack of real-time monitoring and correction mechanisms leads to a decline in user experience and ethical and safety risks.

Method used

By employing reinforcement learning methods, personality deviations are monitored and corrected in real time through personality consistency assessment and closed-loop control. Combined with implicit correction prompts and model fine-tuning, a multi-level optimization mechanism combining correction and reinforcement learning is formed. A personality knowledge base and dialogue memory mechanism are constructed to achieve dynamic perception and progressive optimization.

Benefits of technology

It significantly improves the consistency of personality expression in the big model of roles in long dialogues, maintains the naturalness and stability of the dialogue, reduces the frequency of personality drift and inconsistencies, and enables quantifiable assessment and optimization.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121745210B_ABST
    Figure CN121745210B_ABST
Patent Text Reader

Abstract

This invention relates to the field of natural language processing technology, and discloses a reinforcement learning method for personality consistency alignment and optimization of a large-scale role model, as well as related products. The method involves: acquiring target personality setting information and defining the personality of the large-scale role model; after generating response text in each round of dialogue, combining the target personality setting information with the dialogue context to obtain a personality consistency score for the response text in real time; when the personality consistency score is lower than a scoring threshold, it is processed according to the degree of deviation: for small deviations, implicit personality correction prompts are injected; for large deviations, personality alignment training samples are constructed and added to the training queue; when the triggering condition is met, the personality compatibility reward value is output by the personality reward model as a reward signal, and the model is fine-tuned using a reinforcement learning algorithm with the addition of KL divergence penalty to limit policy drift, forming a closed-loop mechanism of "detection-correction-optimization," effectively suppressing personality drift in long dialogues and significantly improving cross-round personality consistency.
Need to check novelty before this filing date? Find Prior Art