LLM Hacker News 3h ago 2 min read
HN latched onto a pain every heavy coding-tool user knows: the bug is tiny, but the diff balloons anyway. A new write-up turns that annoyance into a measurable benchmark and argues that better prompting and RL can make models edit with more restraint.