EAGER: Using Large Language Models to Model Threats to Sensitive Information

发放政府档案的审查过程可能费时，而且容易出错。大型语言模型可以帮助审查人员确定信息是否已经在公共领域。通过开发一个原型系统和测量不同阶段的性能，该项目旨在估计达到可接受的准确度所需的额外数据和培训。系统的迭代性质和领域专家的参与允许测量和最小化“幻觉”。该项目从知识数据库中挖掘大语言模型的推理能力。它开发了一个语义查询引擎，为准确提取相关信息进行了优化。该项目还采取了一种积极的微调方法，即领域专家训练一个模型，该模型生成查询以从知识库中检索记录，并允许他们通过评估从这些记录中提取的段落来微调检索引擎，然后将其送入大型语言模型进行分析。输出包括通过记录汇编找到的内容的文字说明，并附有记录本身，以供进一步评价和微调。最近发布的记录将作为测试数据，专家将这些信息归类为新的或已知的。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

The review process for releasing government records can be time-consuming and error prone. Large Language Models could help reviewers determine whether information is already in the public domain. By developing a prototype system and measuring performance at different stages, this project aims to estimate the additional data and training required to achieve acceptable levels of accuracy. The iterative nature of the system and the involvement of domain experts allows for measuring and minimizing “hallucination.”The project decouples the reasoning ability of Large Language Models from knowledge databases. It develops a semantic query engine optimized for accurate extraction of relevant information. The project also takes an active approach to fine-tuning, whereby domain experts train a model that generates queries to retrieve records from the knowledgebase, and allows them to fine tune the retrieval engines by assessing the passages that are extracted from these records before they are fed into the Large Language Model for analysis. The output includes text descriptions of what is found through record assembly, accompanied by the records themselves for further evaluation and fine-tuning. Recently released records will serve as test data, with experts categorizing the information as new or already known. Performance metrics are analyzed, considering the impact of data size and composition on accuracy.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.