量子位·2026-05-17 11:42·国内权威源SFT别急着接RL!你的多模态大模型可能一直在“带伤训练”中文摘要在多模态模型进行强化学习前,应先修复SFT阶段的问题,防止“带伤训练”。English SummaryFix SFT issues before moving to RL in multimodal models to prevent training with flaws.原文节选先把SFT挖的坑填了!阅读原文View in English