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

• 类型：arxiv
• 标识：2607.28887
• 链接：https://arxiv.org/abs/2607.28887
• 主分类：evaluation
• 形态：method
• 被引：0
• 被引来源：Semantic Scholar
• S2被引：0
• 影响力被引：0
• TLDR：Teaching deletion during post-training reduces deletion avoidance and improves broader code-editing performance, suggesting the behavior is undertrained rather than beyond reach.
• 待LLM分类：否
• 标题中文：标题 -> 标题中文：增是机器，删是人工：LLM 代码编辑中删除回避的度量与缓解
• TLDR中文：在后训练中教授删除操作可减少删除回避行为并提升更广泛的代码编辑性能，表明该行为是训练不足而非不可达成
• 来源文件：
• /inbox/tom/_candidates/2026-08-05-agent-rag-longcontext-candidates.json
• [S2 enrich]