一句话：提出用岭回归对齐脑活动与预训练模型嵌入，以构建认知启发的 AI 软件开发工具。
本文是一篇立场论文，关注软件工程 AI 工具依赖高维向量嵌入、却未必与开发者目标一致的问题。作者提出将 AI 工具嵌入与人类脑活动翻译得到的嵌入对齐，具体做法是训练岭回归模型，在软件任务中的血氧水平依赖（BOLD）信号与预训练模型嵌入之间建立双向翻译。论文讨论这种翻译模型如何在不同层次支持认知启发的 AI 工具，从高层功能到低层实现细节，并指出从脑活动到向量嵌入的转换、以及在模型训练或微调中融入该信息所面临的开放挑战与潜在解决方案。
AI tools in Software Engineering (SE) rely internally on high dimensional vector embeddings to perform tasks like code generation and debugging. To better coordinate AI tools with developers' goals, we propose to align their embeddings with those translated from human brain activity. This position paper proposes a simple approach for bridging these modalities by training ridge regression models to translate between brain activity, measured as Blood Oxygen Level Dependent (BOLD) signal during software tasks, and pretrained model embeddings. We discuss how such translation models could support cognitively informed AI tools at different hierarchies, from high-level functionality to low-level implementation details. Finally, we highlight open challenges and potential solutions for both translating from brain activity to vector embeddings, and incorporating this information during model training or fine-tuning.