To Add Is Machine, To Delete Is Human: Measuring and Mitigating Deletion Avoidance in LLM Code Editing

This paper studies deletion avoidance in LLM code editing: models often preserve code that the intended patch should remove. Across five leading models on tasks all five solved, deletion recall reached at most 71.7%; models identified the correct file for more than 92% of required deletions but removed the exact line in fewer than 52% of cases. Some 29.0% of passing patches used “Guard-and-Go,” wrapping obsolete code in a guard or fallback. Adding deletion-sensitive tests to 34 SWE-bench Verified tasks reduced four frontier models’ pass rate from 63.2% to 41.9%. The authors introduce CanItDelete, a 200-task deletion-only benchmark.