基于大语言模型的科技计划项目绩效评价研究

Research on Performance Evaluation of Science and Technology Plan Projects Based on Large Language Models

摘要: 针对科技计划项目绩效评价维度单一、过度依赖专家主观经验、非结构化材料处理效率低下，成果真实性与创新性难以开展深度分析等问题，提出一种基于大语言模型的智能绩效评价模型。该模型对项目全周期文本开展深度语义理解，从成果真实性、目标相关性、时间有效性、成果创新性四个维度开展辅助定性研判，输出风险提示与参考线索。以江西省2021~2025年1625项省级科技计划项目为样本开展模型推理，模型共触发风险提示97次，涉及74个项目，完成100个项目的材料筛查、信息比对与风险线索输出平均耗时0.5 h，约为人工开展同等审核工作耗时的十分之一，初步表明该方法在减轻评审工作负担、拓展绩效评价分析维度方面的实用性。

Abstract: Addressing issues such as the single dimension of performance evaluation for science and technology plan projects, over-reliance on subjective experience of experts, inefficient processing of unstructured materials, and insufficient in-depth analysis of the authenticity and innovativeness of outcomes, an intelligent performance evaluation model based on a large language model is proposed. This model conducts deep semantic understanding of project text throughout the entire cycle, completes qualitative evaluation from four dimensions: authenticity of outcomes, relevance to objectives, time effectiveness, and innovativeness of outcomes, and outputs risk warnings and comprehensive research judgments. Using 1625 provincial science and technology plan projects in Jiangxi Province from 2021 to 2025 as test samples, the model triggered risk warnings 97 times, involving 74 projects. Testing showed good consistency between risk warnings and acceptance conclusions; processing 100 projects took an average of 0.5 hours, which is about one-tenth of the traditional review cycle. This preliminarily verifies the practicality of this method in improving evaluation efficiency and deepening analysis dimensions.