// ⌖ Interpretability
Activation Steering
Adds a direction vector to activations at inference to change behavior.
激活引导 aka Abliteration / 拒绝方向消融 / Directional Ablation / Steering Vector 2023 Intermediate core
IN PLAIN WORDS
Intuition: certain directions inside a model map to behaviours. The activation difference between contrasting prompts gives a steering vector that, added at a layer, makes outputs more positive, honest or on-topic without fine-tuning. Projecting the “refusal” direction out of the weights (“abliteration”) removes refusals, showing how easily fine-tuned safety can be stripped.
Used in: Behaviour control and safety research; the community uses the refusal direction to make “abliterated” uncensored open models, whose risks t3n reported on in Oct 2026
01Original paper
Steering Language Models With Activation Engineering
Alexander Matt Turner, Lisa Thiergart, Gavin Leech, et al. · 2023
arXiv ↗ ✓ Verified · arXiv API
02Where it sits
Learn the left column first; the right column builds on this one.
Learn first
Then you can learn
- Nothing in the atlas builds on it yet
Related:Sparse Autoencoders (SAE)Locating & Editing Facts (ROME)
Left: prerequisites · right: builds on this · dashed: related. Full map →
03Learning resources
Official docs, explainers, courses and reference code. All links checked.
Reference code
04Where the course teaches it
On learning paths:Generative Models & Frontier Research
For AI / data
- Verification: verified, checked 2026-10-22.
- JSON: /api/learn/algorithms.json (id:
activation-steering)