The abstract says: "for each model, we find a single direction such that erasing this direction from the model's residual stream activations prevents it from refusing harmful instructions, while adding this direction elicits refusal on even harmless instructions."

The exact words it rests on

Abstract: "for each model, we find a single direction such that erasing this direction from the model's residual stream activations prevents it from refusing harmful instructions, while adding this direction elicits refusal on even harmless instructions."

What the source said when we opened it, on 26 Oct 2026.

The source

Refusal in Language Models Is Mediated by a Single Direction
arXiv (Arditi et al.) · 2024-06-17

Open the sourcearchived copy

Checked

Checked by the notis newsroom on 26 Oct 2026, against the source above.

In the story

How abliteration removes a downloadable model's refusals, and what it costs 26 Oct 2026