Scene-oriented model reconstruction method, system, device and medium

By analyzing the parameter differences of isomorphic models and performing Bayesian optimization, combined with probabilistic surrogate models and Pareto front quality strategies, the problem of low fusion efficiency in large model training is solved, achieving efficient model selection and deployment, meeting the computing resource and performance requirements of different scenarios, and improving system response speed and user experience.

CN122242658APending Publication Date: 2026-06-19HARBIN INSTITUTE OF TECHNOLOGY (SHENZHEN) (INSTITUTE OF SCIENCE AND TECHNOLOGY INNOVATION HARBIN INSTITUTE OF TECHNOLOGY SHENZHEN)
View PDF 0 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
HARBIN INSTITUTE OF TECHNOLOGY (SHENZHEN) (INSTITUTE OF SCIENCE AND TECHNOLOGY INNOVATION HARBIN INSTITUTE OF TECHNOLOGY SHENZHEN)
Filing Date
2026-03-20
Publication Date
2026-06-19

AI Technical Summary

Technical Problem

Existing technologies suffer from high retraining costs and slow iterations in large model training. Coarse-grained fusion lacks diversity and is difficult to adapt to the needs of different budget ranges. Furthermore, fine-grained fusion leads to the curse of dimensionality and low search efficiency, failing to meet the requirements for engineering deployment.

Method used

By constructing a parameter difference analysis mechanism for isomorphic models, importance parameters are extracted. Bayesian optimization and probabilistic surrogate models are used to model in the layered fusion weight space. Combined with the Pareto front quality acquisition strategy, candidate weights are optimized to meet the inference budget constraints and preference strategies of the business task description, thereby achieving the selection and deployment of the optimal candidate model.

Benefits of technology

It significantly improves the search efficiency of model fusion, ensures the exploration of Pareto optimal solution sets within a limited budget, dynamically adapts to computing resource constraints and performance preference requirements, maximizes inference throughput, reduces service costs, and improves system response speed and user experience.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122242658A_ABST
    Figure CN122242658A_ABST
Patent Text Reader

Abstract

This application provides a scenario-oriented method, system, device, medium, and program product for large-scale model reconstruction, relating to the field of large-scale model training. It obtains a business task description and isomorphic first and second models; extracts importance parameters from the layer-by-layer parameter differences between the two models; performs Bayesian optimization in the layer-by-layer fusion weight space, utilizes a probabilistic surrogate model to model the mapping relationship between capability and efficiency, outputs candidate weights with a Pareto front quality-oriented acquisition strategy, continuously improves the front strength and density, and outputs candidate fusion models and offline evaluation records after meeting the required number of evaluations; finally, it selects the optimal model for execution based on hard constraints on inference budget and a preference strategy. This application can maximize inference throughput, reduce service costs, and improve system response speed and user experience while ensuring task accuracy.
Need to check novelty before this filing date? Find Prior Art