Skip to main content
Abliteration reduces refusal behavior in an open-weight LLM by suppressing a refusal-related direction. Abliterated models may answer prompts the source model would refuse, but other capabilities can change too. The name combines ablation and obliteration. Directional ablation can target behaviors beyond refusals; this page focuses on refusal reduction.

​
How it works

Modern instruction-tuned LLMs encode refusal as a recognizable direction in their internal activation space. When a prompt activates that direction strongly, the model produces a refusal (“I can’t help with that”) instead of a substantive answer. Abliteration:
  1. Identifies the refusal direction by computing the difference in mean activations between harmful and harmless prompts on the original model.
  2. Suppresses that direction during inference or edits model weights that contribute to it.
  3. Evaluates refusal behavior and other capabilities after the edit. Results vary by model and method.
Weight-editing methods can modify many values across transformer layers. Reduced refusals do not guarantee unchanged instruction-following, tool use, code generation, or multilingual performance. Abliteration uses a refusal-related direction identified from model activations. Its effect on other behavior must be measured.

​
Origin

Arditi et al. described and demonstrated directional ablation of refusals across multiple model families, including Llama 2, Qwen 1, and Gemma 1. FailSpy published an implementation called abliterator.

​
How abliteration.ai uses this

abliteration.ai hosts abliterated models behind an OpenAI- and Anthropic-compatible HTTP API. Send a request with the OpenAI or Anthropic SDK using https://api.abliteration.ai as the base URL. Refusal behavior varies by model and prompt. The hosted models are exposed as abliterated-model (general-purpose, with vision), abliterated-model-large-v2 (text-only, frontier-scale, GLM-5.3), and abliterated-model-large (text-only, the previous GLM-5.2 large model). See models.

​
What this means for your application

  • Fewer refusals on prompts that fall in the refusal subspace of the original model.
  • No system-prompt jailbreaks needed — refusal reduction is applied to the model rather than added to each prompt.
  • Evaluate other capabilities — abliteration can affect instruction-following, tool use, code generation, and multilingual performance. We run refusal checks and task benchmarks on our hosted models; test your own workload as well.
  • Policies are configurable. Create policies to monitor or block specific request or response behavior. Published policies can apply automatically to projects or only when selected for a run.

​
Further reading

Last modified on September 30, 2026