#unlearning

4 stories · 1 in 7 days · 1 in 30 days

AI安全 2 可解释性 2 arXiv 1 LLM 1 PIR 1 sandbagging 1 对抗攻击 1 批归一化 1

Stories, products and papers tagged "unlearning", newest first.

Nov 4

Papers Builders14:01

Chinese contentarXiv 新论文：用激活碎片化几何检测 LLM 被压制的知识

有人用匿名改写前后的隐状态几何差异，判断模型是真不知道还是憋着不说，代码也开源了，做遗忘评测的可以看看。

#unlearning #LLM #激活分析 #AI安全

arXiv cs.AI@Adar Peleg et al. (4)Source

Sep 24

Papers Builders09:02

Chinese contentHidden not Deleted：神经网络如何抑制纠缠特征

一篇讲特征擦除为什么失败的论文：线性投影擦不干净，遗忘的知识一个标量补丁就能恢复，还解释了 LLM unlearning 复现的机制。

#可解释性 #特征擦除 #unlearning #arXiv

arXiv cs.AI@Akash Samanta et al. (3)Source

Sep 21

Papers Policy14:31

Chinese content给大模型测谎：PIR 从内部状态读出被隐瞒的答案

借法医测谎思路做的 PIR，能从 Gemma、Qwen 等模型内部读出它明明知道却装不知道的答案。

#PIR #sandbagging #unlearning #AI安全

arXiv cs.AI@Hiskias DingetoSource

Sep 11

Papers BuildersSelected09:18

Audit of 263 Batch-Normalized Checkpoints Finds Unlearning Numbers Move Per Checkpoint

This matters to developers because it reveals a potential flaw in how unlearning is verified for these models, which could impact trust in AI systems that rely on unlearning capabilities.

#unlearning #batch normalization #audit #AI safety

arXiv cs.LG@Junlong Shen Xingyu LiSource

unlearning · AI news | AITOP