Visual language model training method and system

By using a multimodal geometric training dataset and a reinforcement learning framework, combined with population relative policy optimization and a dual-track reward function, the spatial intelligence of the visual language model is improved, solving the problems of poor generalization ability and inaccurate evaluation in existing technologies, and achieving more efficient spatial reasoning capabilities.

CN121505632APending Publication Date: 2026-02-10BEIJING ZHONGGUANCUN UNIVERSITY +1
0 Cites 1 Cited by

Patent Information

Application Number
CN202511382359.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-09-25
Publication Date
2026-02-10

AI Technical Summary

Technical Problem

Existing visual language models perform poorly when handling tasks requiring precise spatial perception, exhibiting poor generalization ability, inconsistent training data quality, lack of systematicity, and flawed evaluation and reward mechanisms, thus failing to effectively improve the spatial intelligence of the models.

Method used

We provide a multimodal geometric training dataset, employ a reinforcement learning framework and a population relative policy optimization algorithm, and design a dual-track reward function. Through a high-quality, systematic geometric training dataset and an accurate reward evaluation mechanism, we enhance the spatial reasoning ability of the model.

Benefits of technology

This improved the model's generalization ability on unseen spatial tasks, enhanced the effectiveness and robustness of the training process, and ensured the accuracy and fairness of the evaluation.

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

Abstract

The invention relates to a visual language model training method and system, and the method comprises the steps: providing a multi-modal geometric training data set which comprises a plurality of data units, each data unit comprises digital image data for encoding a geometric figure, a section of text data for describing a question to the geometric figure and a standardized answer encoded in a markup language format supporting machine analysis; generating one or more candidate answers according to the text data and the digital image data in the geometric training data set by applying a reinforcement learning framework and taking the visual language model as a strategy; for each generated candidate answer, a reward function module is applied to allocate a reward value. According to the method, generalized space intelligent improvement is realized, the effectiveness and robustness of training are improved, and the model can be encouraged to explore different problem solving paths and find a correct reasoning chain.
Need to check novelty before this filing date? Find Prior Art