arXiv cs.CL· Hoang-Chau Luong, Lingwei Chen·· 11 小时前AI 评分31
SHIFT：通过外推“收缩-修复”前瞻实现大语言模型遗忘
Unlearning by Extrapolating a Shrink-and-Heal Lookahead
AI 导读
SHIFT 通过将微调后模型向预微调权重收缩并短暂重学保留集，估计指向重训练模型的遗忘方向，再外推为蒸馏目标迭代优化。在 OpenUnlearning 统一协议下对比 11 个基线，SHIFT 在全部 TOFU 设置中取得最高综合得分，与 oracle 的对齐程度平均比次优方法近 3.7 倍；在 MUSE 上保留知识最多。
正文 · 原文
Abstract:Unlearning aims to make a large language model behave like one retrained without its forget set. Prevailing likelihood-suppression objectives, however, lack an explicit forgetting target: they continually reward lower forget-set likelihood, making forgetting sensitive to the optimization budget. We address this limitation by estimating the direction toward the oracle without training it. Our key idea is to shrink the fine-tuned model toward its pre-fine-tuning weights and briefly re-learn the retain set, thereby removing the forget-specific contribution; in the linearized regime, this shrink-and-heal lookahead moves the model a fixed fraction of the way to the retrained model, its unique fixed point. Our method, SHIFT, extrapolates the lookahead's per-token changes into distillation targets and iteratively refines them. Evaluated against eleven baselines under a unified OpenUnlearning protocol, SHIFT achieves the highest aggregate score across all TOFU settings and the closest oracle alignment, 3.7x closer on average than the next-best method; on MUSE, it preserves the most retain knowledge. Under larger optimization budgets, baselines overshoot the oracle, while SHIFT remains near it, even under relearning attacks.
| Subjects: | Machine Learning (cs.LG); Computation and Language (cs.CL) |
| Cite as: | arXiv:2611.00374 [cs.LG] |
| (or arXiv:2611.00374v1 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2611.00374 arXiv-issued DOI via DataCite (pending registration) |
Submission history
From: Hoang-Chau Luong [view email]
[v1]
Fri, 30 Oct 2026 20:57:41 UTC (50 KB)
来源：arXiv cs.CL · arxiv.org