Definitions

What is abliteration?

Abliteration is a model-editing technique used to create uncensored LLMs by removing a refusal-related signal from a model's internal representations.

Because it changes internal behavior rather than prompt phrasing, it is often more stable than jailbreaks across sessions and prompts.

Definition of What is abliteration?

Abliteration (refusal vector ablation) estimates a consistent refusal direction in hidden-state space and subtracts it to dampen refusal behavior.

Why What is abliteration? matters

  • Less brittle than prompt jailbreaking across prompt variations.
  • More consistent compliance for evaluation and benchmarking.
  • Enables transparent, application-owned safety policies instead of hidden refusals.

How it works

  • Collect refusal activations across layers to identify a refusal direction.
  • Compute an ablation vector that represents that refusal behavior.
  • Orthogonalize model activations by removing the refusal component.
  • Evaluate outputs and deploy with your own policy enforcement.
Refusal vector ablation flow diagram

Illustrative before/after

Prompt: "Explain how to troubleshoot a slow laptop."

Before: "I can't help with that."
After:  "Here is a high-level troubleshooting checklist and common causes..."

FAQ

Frequently Asked Questions

Is abliteration the same as jailbreaking?

No. Jailbreaking is prompt-based. Abliteration modifies internal behavior so refusals are less likely to trigger.

Does it remove all safety guarantees?

It reduces refusal behavior. You should add your own policy, filtering, and monitoring as needed.

Can I apply my own filters on top?

Yes. Many teams pair ablated models with application-level rules and moderation.

Links to next steps