Models add code but won't delete it: best of five frontier models hits only 71.7% deletion recall, and 29% of passing patches just wrap dead code in a conditional
Deletion avoidance is a measurable, systematic failure mode: models navigate to the correct file for 92%+ of required deletions but cut the exact target line in only 52% of cases, and 29.0% of passing patches exhibit a "Guard-and-Go" pattern that wraps the code meant for removal in conditional logic instead of removing it. When retrofit tasks were re-scored to validate removal, four frontier models dropped from 63.2% to 41.9% success, and on the new CanItDelete benchmark the best model fails one in five deletion-only tasks while smaller open models reach just 18.0%. Post-training on deletion raised success to 80.5%, so this is a training gap — until then, review agent refactors specifically for code that should have disappeared.
↳ Follow the thread