LLMs struggle to delete code, hindering maintainability, new research finds

A new research paper identifies a significant issue in large language models' code editing capabilities: deletion avoidance. Models tend to retain code that should be removed, often by wrapping it in guards or fallbacks instead of deleting it. This behavior, observed across leading models on the SWE-bench benchmark, leads to code that is harder to maintain, even when tests pass. The research also introduces a new benchmark, CanItDelete, to specifically test deletion capabilities and suggests that post-training interventions could help mitigate this problem. AI

IMPACT Identifies a key limitation in LLM code editing that could impact software development workflows and suggests potential solutions.

RANK_REASON Research paper published on arXiv detailing a specific limitation in LLM code editing capabilities. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Amir M. Ebrahimi
• CanItDelete
• GPT 5.6 "Sol"
• Guard-and-Go
• large-language models
• SWE-bench

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →