Abliteration
Abliteration is a technique that stops a language model from refusing requests by finding the single internal direction that carries refusal and editing the weights so the model can no longer write to it, without any retraining.
Also called: refusal ablation, refusal direction removal, abliterated model
By Divya Sharma · Checked Nov 16, 2026 · How we review
Chat models are trained to follow instructions and to refuse harmful ones. In 2024 Arditi and colleagues showed that, across 13 open chat models up to 72B parameters, this refusal runs through one direction in the model's internal activations. Erase that direction and the model stops refusing; add it and the model refuses even harmless requests. Abliteration is the practical recipe built on that finding.
How it works
You run the model on two sets of prompts, one harmful and one harmless, and record the activations at the last token. The difference between the two averages gives a candidate refusal direction for each layer, and you keep the one that works best. Then you orthogonalise the matrices that write into the residual stream against that direction, so the model weights can no longer produce it. The result is an ordinary checkpoint that can be shared like any other open-weight model.
Abliteration vs fine-tuning
Fine-tuning changes behaviour by training on examples with gradient updates. Abliteration needs only the two prompt sets and no training: it is a one-off edit of existing weights that removes a behaviour rather than teaching a new one. It is a white-box method, so it only works on models whose weights you hold.
Costs to quality
The paper reports little effect on other capabilities, but results vary by model. In Labonne's walkthrough, abliterating the 8B model Daredevil-8B lowered its benchmark scores. A light round of DPO preference training afterwards recovered most of the drop, while he warned that more supervised fine-tuning would likely break the model.
Risks
Abliteration removes refusals for every request, not only the ones a user finds over-cautious. Arditi and colleagues describe it as a white-box jailbreak and say it shows how brittle safety fine-tuning is. An abliterated model has no built-in refusal, so any limits on what it produces must come from outside the model, through guardrails and testing such as red teaming. It also means safety training on a released model can be undone by anyone who has the weights.
Related terms
- Jailbreak (LLM)
- A jailbreak is a prompt or technique that gets a language model to produce output its safety training is meant to refuse, such as harmful instructions, by role-play, obfuscation, odd encodings or long adversarial text.
- Model weights
- Model weights are the learned numbers inside a neural network, set during training, that decide how it turns input into output. A checkpoint is a saved file or set of files containing those weights.
- Open-weight model
- An open-weight model is a model whose trained weights are published for anyone to download and run, often without the training code or full training data, so it is not necessarily open source.
- AI alignment
- AI alignment is the work of making an AI system's behaviour match what its users and developers actually intend, so it is helpful, honest and avoids harm, rather than just predicting likely text or optimising a narrow score.
- AI red teaming
- AI red teaming is the practice of deliberately attacking a model or AI application before and after release, with human testers or automated tools, to find harmful outputs, jailbreaks, injections and other failures so they can be fixed.
- Guardrails (AI)
- Guardrails are runtime checks placed around a language model that inspect inputs and outputs and block, rewrite or redirect them, so an application stays on topic, avoids harmful content and follows rules the model alone would not reliably keep.