Web IR NLP Group @ NUS
Web IR NLP Group @ NUS
Tour
News
People
Events
Gallery
Publications
Projects
Contact
Light
Dark
Automatic
Code Repair
LLMs for Code: Faithful and Minimal Code Editing
We study whether LLM code repairs stay minimal and faithful to the original program, rather than only whether they pass the tests, and how models can be trained to edit code minimally.
Tongyao Zhu
,
Wei Hern Lim
,
Min-Yen Kan
Paper
Code & Data
When Models Edit Too Much: On the Fidelity of Minimal Code Edits
LLM code repairs often pass the tests while rewriting far more code than needed. We measure this over-editing on a benchmark with known minimal patches and show that reinforcement learning can teach models to edit minimally.
Tongyao Zhu
,
Wei Hern Lim
,
Min-Yen Kan
PDF
Cite
Code
Project
DOI
ACL Anthology
arXiv
Cite
×