拒绝方向消融(Abliteration)原理详解:从 Arditi 论文到 OBLITERATUS
拒绝方向消融原理详解:Arditi 论文的单方向发现、权重正交化怎么做、KL 散度的含义、副作用,以及与微调、DPO、提示词越狱的对比。
拒绝方向消融原理详解:Arditi 论文的单方向发现、权重正交化怎么做、KL 散度的含义、副作用,以及与微调、DPO、提示词越狱的对比。
DeepSeek V4.1 Flash 1017 Obliterated explained: refusals from 81.62% to 4.30%, flat MMLU, the limits of self-reported evals, a 1.51 TB checkpoint and the risks.
Refusal direction ablation explained: the single-direction finding, weight orthogonalization, what KL divergence means, side effects and comparisons.