Research on active retrieval enhancement for large language models based on multi-model integration
Year: 2025
Type: conference-paper
Abstract: Large language models have limitations when dealing with long-tail knowledge and specific scenarios. To improve performance, retrieval-augmented language models have emerged, but traditional methods are inefficient and prone to excessive retrieval. This paper combines ERNIE-3.5 and GPT-3.5-Turbo to propose an active retrieval augmentation model aimed at improving zero-shot learning performance. By introducing an active retrieval mechanism and specific prompt templates to stimulate the model’s self-reflection ability, it effectively addresses the lack of model performance in specific scenarios. Experimental results show that the framework achieves an average accuracy of 66.10% on the MMLU benchmark and performs well in zero-shot learning. Ablation experiments demonstrate the effectiveness of the method, reducing retrieval time and improving accuracy. In the future, we will address the issues of insufficient knowledge base information and model self-reflection ability, and explore more application scenarios.
DOI: 10.1117/12.3062366
Authors: Baosheng Yin, Zhenlai Xu
Institution: Shenyang Aerospace University
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4409784734