Yongcheng Huang.

CoReFusion treats code refactoring as a denoising task — viewing suboptimal variable names as "noise" to be removed. Using diffusion language models, we show they excel at filling masked positions but struggle to localize which identifiers need renaming without guidance. This "filling vs finding" gap reveals key insights into how dLLMs reason about code.

Dfusion combines two pretrained diffusion models (RDiffusion and DiT) through a lightweight fusion CNN for raindrop removal. Each model excels at different aspects — one for smooth textures, the other for fine details. The fusion network merges their strengths, achieving robust restoration in both day and night conditions.

We evaluate code LLMs in non-English contexts through an open-coding study of 12,500 labeled generations across five languages (Chinese, Dutch, English, Greek, Polish). Our analysis yields a taxonomy of 26 error categories, revealing that while models often produce partially correct comments, standard neural metrics fail to reliably distinguish meaningful completions from noise.

After the emergence of BERT, Large Language Models have demonstrated remarkable multilingual capabilities. However, current evaluations primarily focus on English use cases. This study assesses LLM performance in generating Chinese Java code comments through open coding, revealing the prevalence of semantic errors (especially hallucinations) while showing an absence of grammatical issues due to Chinese language characteristics. We also validated the potential for quantitatively analyzing semantic errors using cosine similarity of word embeddings.

This research proposes a shift to a structured, grammar-based methodology for measuring linguistic coverage of code-generating LLMs. By using formal grammars as a definitive checklist, we analyze code corpora to create linguistic coverage reports, revealing representational imbalances where a few simple constructs dominate while complex features exist in a long tail. We also quantify the "tokenization offset" between grammar-based and token-based decompositions.