Molecular multi-modal large language model construction method, reasoning method and device based on parameter space alignment

By using parameter space alignment, and by injecting molecular structural information into the large language model using graph neural networks and adaptive weight generators, the problems of high computational complexity and fragmented structural information in existing technologies are solved, and efficient and accurate molecular multimodal large language model inference is achieved.

CN122290793APending Publication Date: 2026-06-26CHONGQING UNIV
1 Cites 0 Cited by

Patent Information

Application Number
CN202610402999.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-03-30
Publication Date
2026-06-26

AI Technical Summary

Technical Problem

Existing molecular multimodal large language models suffer from high computational complexity, fragmented structural information, and severe context window contamination when dealing with complex molecular structures, resulting in high inference and training costs and making it difficult to effectively handle scientific tasks such as multi-step synthetic path planning or reaction mechanism analysis.

Method used

A parameter space alignment-based approach is adopted, which uses a graph neural network encoder to extract node-level feature representations of molecular structure diagrams, and obtains low-rank weight update amounts of a large language model through an adaptive weight generator, which are then injected into the parameter space of the large language model to achieve deep perception and efficient training of molecular structures.

Benefits of technology

It enhances the model's overall perception of complex molecular topology, reduces computational overhead and training costs, improves inference accuracy and efficiency, avoids fragmentation of structural information, and enhances the model's performance in scientific tasks.

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

Abstract

The invention belongs to the technical field of artificial intelligence and chemical informatics crossing, and provides a parameter space alignment-based molecular multi-modal large language model construction method, reasoning method and device, and the method comprises the steps: constructing an instance set; constructing a molecular multi-modal large language model, wherein the molecular multi-modal large language model comprises an adaptive weight generator, a graph neural network encoder and a large language model; using the instance set to train a molecular multi-modal large language model to update parameters of an adaptive weight generator; a graph neural network encoder extracts node-level feature representation of the molecular structure graph; the adaptive weight generator obtains a low-rank weight update quantity of a target weight component in the large language model based on the node-level feature representation, and injects the low-rank weight update quantity into an original weight of the target weight component; and the large language model after injection of the low-rank weight update quantity generates an actual text response based on an input training text instruction. The molecular structure is integrated to a large language model parameter layer, and depth perception of the molecular structure is realized under the condition that the length of an input sequence is not changed.
Need to check novelty before this filing date? Find Prior Art