CodeT5+
PulseAugur coverage of CodeT5+ — every cluster mentioning CodeT5+ across labs, papers, and developer communities, ranked by signal.
-
DocSync agent uses code structure and LLMs to maintain software documentation
Researchers have developed DocSync, an agentic system designed to automatically maintain software documentation by ensuring it remains consistent with evolving code. The system uses Abstract Syntax Trees and Retrieval-A…
-
考古团队微调代码模型以检测AI生成代码
考古团队的研究人员开发了一个用于检测AI生成代码的系统，并参加了SemEval-2026 Task 13。他们的方法包括微调几个预训练的代码模型，如CodeBERT和CodeT5+，采用不同的策略进行AI生成代码的二元分类和生成模型的多元归属。他们的最佳提交取得了有竞争力的结果，在二元分类中排名第6位（共81支队伍），在模型归属中排名第7位（共34支队伍）。
-
Language models improve static program slicing with dataflow pretraining
Researchers have developed Sliceformer, a new method for static program slicing that uses language models to identify relevant code segments. This approach addresses limitations in existing learning-based techniques by …