H2DF: Calibrated Target Bands for Controllable Language-Model Unlearning

查看原文 / View Original 编辑 / Edit

Computer Science > Computation and Language arXiv:2610.26864 (cs) [Submitted on 26 Jun 2026] Title:H2DF: Calibrated Target Bands for Controllable Language-Model Unlearning Authors:Jupiter Yao View a PDF of the paper titled H2DF: Calibrated Target Bands for Controllable Language-Model Unlearning, by Jupiter Yao View PDF HTML (experimental)

计算机科学 > 计算与语言 arXiv:2610.26864 (cs) [提交于 2026 年 6 月 26 日] 标题：H2DF：用于可控语言模型遗忘的校准目标带作者：Jupiter Yao 查看标题为《H2DF：用于可控语言模型遗忘的校准目标带》的论文 PDF，作者：Jupiter Yao 查看 PDF HTML（实验性）

Abstract:Gradient ascent (GA) is a simple language-model unlearning baseline, but it continues increasing forget loss after useful forgetting has occurred. This unbounded pressure can damage retained utility and make forgotten examples abnormally high-loss. We introduce H2DF, a calibrated target-band objective that applies GA-like pressure below a data-dependent lower target and suppresses, or reverses, that pressure after the target region is reached.

摘要：梯度上升（GA）是一种简单的语言模型遗忘基准，但在实现有效遗忘后，它会持续增加遗忘损失。这种无限制的压力可能会损害保留的效用，并使被遗忘的样本产生异常高的损失。我们引入了 H2DF，这是一种校准目标带目标函数，它在数据依赖的下限目标之下施加类似 GA 的压力，并在达到目标区域后抑制或反转该压力。

The targets are estimated from matched non-member examples before training, so the base objective requires one forget-batch forward/backward pass per optimization step; our reported TOFU variant adds sparse retain replay every tenth step. We instantiate a one-sided objective for TOFU question answering and a two-sided objective for MUSE-News language modeling.

目标是在训练前根据匹配的非成员样本估算的，因此基础目标函数在每个优化步骤中需要一次遗忘批次的前向/反向传播；我们报告的 TOFU 变体在每十个步骤中增加一次稀疏的保留重放。我们为 TOFU 问答实例化了一个单侧目标函数，并为 MUSE-News 语言建模实例化了一个双侧目标函数。

Across three TOFU seeds, H2DF obtains a forget NLL of 1.528 with margin 0.049 and retain NLL of 1.165 with margin 0.047, matching GradDiff while using retain replay only every tenth step and showing substantially lower variability than GA. On MUSE-News, H2DF improves retention over GA and NPO and is competitive with SimNPO, but does not dominate it.

在三个 TOFU 种子实验中，H2DF 获得的遗忘 NLL 为 1.528（边际 0.049），保留 NLL 为 1.165（边际 0.047），在仅每十步使用一次保留重放的情况下与 GradDiff 持平，且表现出比 GA 低得多的变异性。在 MUSE-News 上，H2DF 在保留效果上优于 GA 和 NPO，并与 SimNPO 具有竞争力，但并未占据绝对优势。

Ablations show that the upper target reduces overshoot and that validation-based checkpoint selection is important. These results support calibrated target bands as a practical mechanism for controlling forgetting strength, while also exposing sensitivity to random seed and calibration quality.

消融实验表明，上限目标减少了过冲，且基于验证的检查点选择非常重要。这些结果支持将校准目标带作为控制遗忘强度的实用机制，同时也揭示了其对随机种子和校准质量的敏感性。