LIFT the Veil for the Truth: Principal Weights Emerge after Rank Reduction for Reasoning-Focused Supervised Fine-Tuning
揭开真相的面纱:秩减少后主权重的浮现使以推理为导向的监督微调得以提升
机构 * University of California, Berkeley, CA, USA(加州大学伯克利分校) ; International Computer Science Institute, CA, USA(国际计算机科学研究所) ; Lawrence Berkeley National Laboratory, CA, USA(伯克利国家实验室) ; Dartmouth College, NH, USA(达特茅斯学院) ; University of Exeter, Exeter, UK(埃克塞特大学) ; University of Oxford, Oxford, UK(牛津大学) ; University of Surrey, Guildford, UK(萨里大学) ; Tsinghua University, China(清华大学) ; Eindhoven University of Technology, the Netherlands(埃因霍温理工大学)
专题命中 数学推理 :reasoning(title,abstract);分类 cs.CL、cs.AI、cs.LG
AI总结 本文提出LIFT方法,通过秩减少后选取主权重进行微调,提升推理能力,同时保持内存效率,优于全微调和LoRA。
Comments ICML 2025