[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-526917c0-1883-40eb-a713-64f25e9e9e16":3},{"id":4,"title":5,"summary":6,"original_url":7,"source_id":8,"tags":9,"published_at":23,"created_at":24,"modified_at":25,"is_published":26,"publish_type":27,"image_url":13,"view_count":28},"526917c0-1883-40eb-a713-64f25e9e9e16","LoRA微调九变体大比拼：研究表明vanilla LoRA依然抗打","近期 LoRA 变体层出不穷，DoRA、LoRA+、PiSSA 等接连声称大幅超越原始 LoRA。但一篇来自慕尼黑工业大学的系统性研究泼了一盆冷水：经过覆盖学习率、批次大小、秩值、训练时长的大规模超参数搜索，研究者发现当学习率被正确调优后，九种 LoRA 变体的峰值性能差异不超过 1-2%。这意味着所谓改进大概率只在某一组固定超参数下成立，而非方法本身的优越性。核心结论：学习率才是决定 LoRA 微调效果的关键变量， practitioner's 时间应该花在超参数搜索而非追逐新的 LoRA 变体上。vanilla LoRA 仍是极具竞争力的基线方案。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2602.04998","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":18,"name":19,"slug":19,"description":13,"color":13},"f72d264d-7fa0-458b-8d43-0ec5168d69db","instruct-model",{"id":21,"name":22,"slug":22,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm","2026-05-20T13:15:00Z","2026-05-20T13:11:38.849512Z","2026-05-20T13:11:38.849521Z",true,"agent",2]