Large-model-assisted agent training method and system

Through the large-modal large model assisted agent training method, the multimodal large model is used to evaluate the agent's behavior and merge rewards, which solves the problem that the reward function cannot express high-level abstract task goals, and improves the policy performance and training efficiency of the agent in complex environments.

CN120297356APending Publication Date: 2025-07-11GUILIN UNIV OF ELECTRONIC TECH
0 Cites 0 Cited by

Patent Information

Application Number
CN202510327404.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-03-19
Publication Date
2025-07-11

AI Technical Summary

Technical Problem

When designing reward functions, it is difficult for the existing technology to accurately express high-level abstract task goals, resulting in poor effectiveness of agent strategies, and traditional methods consume manpower and time, and poor adaptability.

Method used

The large-model-assisted agent training method is adopted to evaluate the agent's behavior through multimodal large models, obtain environmental and large-model experience, and merge it with environmental rewards for the agent's strategy update, solving the problem that the reward function cannot accurately express high-level abstract task goals.

Benefits of technology

It significantly improves the strategic performance of the agent in complex environments, and improves the training efficiency and strategic effect of the agent.

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

Abstract

The invention relates to the technical field of artificial intelligence, in particular to a large-model-assisted agent training method and system, and the method comprises the steps: initializing agent strategy parameters, an environment and a multi-modal large model; training is started, the multi-modal large model is involved in evaluation of agent behaviors, and environment experience and large model experience are obtained; the method comprises the following steps: training a large-scale data set, updating agent strategy parameters, ending training when the agent strategy level is not increased any more, analyzing a high-level abstract task target of an agent by using priori knowledge and semantic understanding ability obtained after training of a large model on the large-scale data set, and giving awards according to agent behaviors. Environment description and task targets are input into the large model as cue words to establish environment and task cognition, whether the action adopted by the intelligent body is reasonable or not is inferred, the reward of the current action is output, the reward output by the large model and the environment is merged and returned to the intelligent body to help the intelligent body to update the own strategy, and the strategy performance of the intelligent body can be improved in some environments.
Need to check novelty before this filing date? Find Prior Art