This paper introduces Late Code Chunking (LC²), a chunking strategy designed to improve the semantic understanding of code segments for Large Language Models (LLMs). Repository-level code completion requires predicting the completion of unfinished code by leveraging cross-file context spread across a repository. However, when retrieved fragments have missing semantics—the loss of structural or behavioral information during chunking—LLMs struggle to interpret the target code. To address this, LC² refines retrieved chunks by constructing a dual context: a "Code Retrieval Context" optimized for similarity-based search, and a "Code Comprehension Context" that serves as a late enrichment step through context expansion and augmentation. This dual-context design reduces information loss due to chunking and enhances the ability of LLMs to utilize retrieved code. Additionally, we introduce an Asymmetric Query-Chunk Sizing strategy to further optimize retrieval quality by minimizing query noise. Our experiments demonstrate that LC² provides robust performance gains, achieving a statistically significant 19.7% improvement in Exact Match accuracy on the CrossCodeEval benchmark compared to the best existing chunking method.

Research on automated program repairs using transformer-based models has recently gained considerable attention. The comprehension of the erroneous behavior of a model enables the identification of its inherent capacity and provides insights for improvement. However, the current landscape of research on program repair models lacks an investigation of their false behavior. Thus, we propose a methodology for diagnosing and treating the false behaviors of transformer-based program repair models. Specifically, we propose 1) a behavior vector that quantifies the behavior of the model when it generates an output, 2) a behavior discriminator (BeDisc) that identifies false behaviors, and 3) two methods for false behavior treatment. Through a large-scale experiment on 55,562 instances employing four datasets and three models, the BeDisc exhibited a balanced accuracy of 86.6% for false behavior classification. The first treatment, namely, early abortion, successfully eliminated 60.4% of false behavior while preserving 97.4% repair accuracy. Furthermore, the second treatment, namely, masked bypassing, resulted in an average improvement of 40.5% in the top-1 repair accuracy. These experimental results demonstrated the importance of investigating false behaviors in program repair models.