Primary sourcearXiv: DeepSeek
View sourceEvent ·Official first-hand
一种无需重新训练的模型知识移除方法
本文提出了一种名为ARIA的新方法,它通过稀疏自编码器在测试时移除大语言模型中的特定知识,而不修改模型权重。该方法在DeepSeek-R1和Gemma-3等模型上进行了测试,在TOFU、R-TOFU和WMDP等基准测试中,成功减少了WMDP-cyber基准的遗忘集准确率,同时保持了MMLU基准的得分仅下降1%以内。
Current read
这个方法挺有意思,不用改模型权重就能移除特定知识,在DeepSeek和Gemma上效果还不错,感兴趣的可以看看。
3 sources44° AI heatLast updated 2026/9/14 18:58:09
Evidence
3 sourcesConflict check
Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.