← 研踪 PaperTrailarXiv:2611.10881 · v1
Robust and Efficient Continual Model Merging via Global Singular Subspace Separation and Restoration
首次提交 2026-11-08 02:05
版本更新 2026-11-08 02:05
雷达观测 2026-11-10 11:14
摘要
Continual model merging incrementally integrates task-specific models without revisiting past models or data. However, repeated merging entangles past task components and distorts their singular geometry, making balanced task-wise allocation increasingly difficult. We propose Global Singular Subspace Separation and Restoration (GS3R), which augments the merged update with a compact restoration state. At each merge, GS3R restores the retained singular geometry, separates task-wise subspaces, and globally reallocates the available rank across tasks fairly. To reduce merging-time memory overhead, we further introduce Pre-merge Restoration and Flushing (PRF). Experiments on vision, NLP, and multimodal benchmarks show that GS3R provides a strong accuracy-efficiency trade-off.
作者备注：NeurIPS 2026. Code: https://github.com/ewsn1593/GS3R
观测到的版本
相邻记录可能跳过中间版本；首次发现 v2 及以上也会标为版本更新。
v1 · 2026-11-08 02:05 · 最新观测版本
Robust and Efficient Continual Model Merging via Global Singular Subspace Separation and Restoration
Continual model merging incrementally integrates task-specific models without revisiting past models or data. However, repeated merging entangles past task components and distorts their singular geometry, making balanced task-wise allocation increasingly difficult. We propose Global Singular Subspace Separation and Restoration (GS3R), which augments the merged update with a compact restoration state. At each merge, GS3R restores the retained singular geometry, separates task-wise subspaces, and globally reallocates the available rank across tasks fairly. To reduce merging-time memory overhead, we further introduce Pre-merge Restoration and Flushing (PRF). Experiments on vision, NLP, and multimodal benchmarks show that GS3R provides a strong accuracy-efficiency trade-off.
查看此版本解读 →查看此版本原文 ↗