杨贵妃传媒

This research project investigates the development of an autonomous, iterative pipeline designed to enhance the capabilities of small- to mid-sized large language models (LLMs) by enabling them to refine both system and user prompts for a given task. We explore strategies for self-prompting frameworks in which the LLM iteratively analyzes a user prompt, generates optimized system-level instructions, and refines the prompt to improve task-specific performance in terms of quality and relevance of the generated response.

The largest state-of-the-art LLMs often demonstrate superior capabilities. However, smaller models are more accessible for resource-constrained environments, such as edge and personal devices. Enhancing the utility of these lightweight LLMs without fine-tuning or expanding their parameter counts could significantly broaden their practical applications.

To assess the effectiveness of this self-prompting framework, the project will establish both qualitative and quantitative benchmarks, including NLP-based similarity scoring and task-specific performance metrics. The interdisciplinary scope will be supported through collaboration with subject matter experts across multiple academic fields to design diverse input prompts and evaluate the quality of the generated outputs.

While this project aims to expand the capabilities of resource-efficient LLMs, we also seek to identify reusable, LLM-specific strategies for improving prompt effectiveness across domains. The findings may later be integrated into existing self-correcting LLM pipelines for tasks involving scientific code generation, contributing to the further advancement of autonomous LLM techniques in scientific contexts.