发表机构
Tsinghua University; DiDi Autonomous Driving（清华大学; 滴滴自动驾驶）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
提出事后JANUS修正框架，通过参数空间正交性投影恢复历史知识，结合多步自适应机制，缓解微调中的稳定性-可塑性困境。
AI 中文摘要
在新任务上微调基础模型不可避免地会遭受灾难性遗忘。尽管现有工作尝试在参数高效微调方法的基础上缓解这一问题，但它们采用了过度受限的子空间正交性条件。本文提出了一种纯粹事后且与微调方法无关的权重修正框架，实现了参数空间正交性，这是保持历史性能到一阶精度的必要且充分条件。通过将参数更新投影到雅可比零空间（JANUS）中，我们的方法显著恢复了受损的历史知识，且不干扰底层微调过程。为克服雅可比近似的局部有效性，我们进一步提出了一种多步自适应修正机制，利用JANUS位移动态验证有效信任区域并调整步长。结合我们提出的幽灵投影、幽灵方向比较和序列级奇异值分解压缩技术，JANUS还实现了极佳的时间和空间效率。实验表明，JANUS能无缝集成各种微调方法，通过恢复历史知识同时保持下游任务适应性，显著缓解稳定性-可塑性困境。
英文摘要
Fine-tuning foundation models on new tasks inevitably suffer from catastrophic forgetting. While existing works attempt to mitigate this on the basis of parameter-efficient fine-tuning methods, they adopted an overly restrictive Subspace Orthogonality condition. In this paper, we introduce a purely post-hoc and tuning-agnostic weight rectification framework that achieves Parameter Space Orthogonality, which is the necessary and sufficient condition for preserving historical performance to the first order. By projecting parameter updates into the JAcobian NUll Space (JANUS), our method significantly recovers compromised historical knowledge without interfering with the underlying fine-tuning process. To overcome the local validity of the Jacobian approximation, we further propose a Multi-step Adaptive Rectification mechanism that utilizes the JANUS shift to dynamically verify the valid trust region and adjust step sizes. Coupled with our proposed ghost projection, ghost orientation comparison, and sequence-level singular value decomposition compression techniques, JANUS also achieves great temporal and spatial efficiency. Experiments demonstrate that JANUS seamlessly integrates with various fine-tuning methods, significantly mitigating the stability-plasticity dilemma by recovering historical knowledge while preserving downstream task adaptation.
