Xijia Tao, Ziru Liu, Shansan Gong, Jiacheng Ye, Kecheng Chen, Zirui Wu, Lin Zheng, Xinyu Fu, Rui Liu, Lingpeng Kong
Author
Xijia Tao, Ziru Liu, Shansan Gong, Jiacheng Ye, Kecheng Chen, Zirui Wu, Lin Zheng, Xinyu Fu, Rui Liu, Lingpeng Kong
Date
Key takeaways · AI-distilled
The study compares four editing interfaces for masked diffusion language models on CanItEdit: whole-file rewriting, search-and-replace, locate-then-infill, and tokenThe chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.Full definition →-level editing.
Diffusion models produce coordinated changes well when the correct edit locations are supplied, but lose much of that ability when they must predict the locations themselves, a gap the authors call the composition gap.
Giving the model the intact original code helps it fill multiple edit regions, but does not solve the problem of choosing which regions to edit.
A sentence-level Wiki editing probe shows the same gap between supplied and predicted locations, suggesting the problem is not specific to code.
Terms in this piece · Glossary
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
Why it matters
If you consider diffusion models for code editing, the bottleneck is choosing edit regions, not generating the changes. Regions must cover every required change yet stay tight, because widening them to ensure coverage sharply reduces success.
articleDiffs vs. Whole Files: An Empirical Comparison of Iterative Edit-Based and Direct Generation for Flutter/Dart Code ModelsAndrej Andrejev
articleTalking to Itself While Coding: What Makes Comments Help Code Generation?Dangfeng Pan, Zhensu Sun, Cenyuan Zhang, David Lo, Xiaoning Du
articleDon't Let the Model Write the YAML: Deterministic, Minimal-Diff GitOps Remediation from LLM-Proposed Field ChangesPruthvi Davineni