Intel / article

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

Source: arxiv.org
Author: Amir M. Ebrahimi, Mohammed Mehedi Hasan, Aaditya Bhatia, Gopi Krishnan Rajbahadur, Ahmed E. Hassan
Date: Published Aug 3, 2026

Context

A coding can pass a repair test while retaining code the intended edit removes. The deletion-avoidance study's abstract reports that 29% of passing patches surrounded targeted code with conditions or fallbacks. To test whether ordinary evaluations overlook this behavior, the researchers added checks to 34 tasks that failed when the targeted code remained. Across four models, the reported pass rate fell from 63.2% to 41.9%. The added removal checks exposed a gap between passing existing behavioral checks and completing the requested removal; they did not show that the models themselves had become worse.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
SWE-bench — The standard benchmark for AI coding agents: real GitHub issues from real repositories, scored by whether the agent's patch passes the project's own tests.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.

Key quotes

“deletion recall against the developer patch reaches at most 71.7% even on tasks all five solve”

“29.0% of passing patches wrap the targeted code in a guard or fallback, a pattern we call Guard-and-Go”

“four frontier models spanning closed and open weights fall from 63.2% to 41.9%”

“teaching deletion during post-training reduces deletion avoidance and improves broader code-editing performance, suggesting the behavior is undertrained rather than beyond reach”

Read the source arxiv.org