CogAdapt：认知信号引导的代码大模型稀疏适配
原标题：CogAdapt: Cognition-informed Sparse Adaptation of Code LLMs
AI 摘要
研究提出 CogAdapt 框架，利用人类阅读代码时的脑电（EEG）和眼动注意力数据学习程序级与 token 级先验，再结合冻结模型对任务的响应，决定微调时更新哪些 transformer 块及分配多少适配量。在 Qwen 和 GLM 上，该方法在 LiveCodeBench 和 BigCodeBench 上取得最佳 pass@1，相比常规微调在 LiveCodeBench 上提升 10.86 和 6.29 个百分点，同时将可梯度更新的适配参数减少 86.21–87.21%。
正文节选
CogAdapt: Cognition-informed Sparse Adaptation of Code LLMs Abstract. Large language models (LLMs) have become increasingly capable of generating code. However, achieving stronger code-generation performance still often relies on costly model adaptation, i.e., fine-tuning pretrained model parameters. Meanwhile, prior studies have shown meaningful correspondence between how humans process code and how neural models allocate attention or internal computation. Building on this relationship, recent