Heretic overview: automated model refusal removal

Heretic is an open source tool for modifying refusal behavior in open-weight AI models. The speaker describes it as an automated approach to directional ablation, also called abliteration, that edits model weights without fine-tuning. This is an explanation of the technique and reported results rather than an installation tutorial.

The explanation draws on a 2024 paper linking refusal behavior to a direction in a model's residual stream. According to the speaker, Heretic estimates that direction from the difference between harmful and harmless prompt activations. It then uses Optuna to search for ablation parameters that reduce refusals while limiting KL divergence from the original model.

For Gemma 3 with 12 billion parameters, the speaker reports refusal removal matching a manual ablation method, with KL divergence of 0.16 versus 1.04. The speaker interprets the lower value as less damage to reasoning; the transcript does not establish broader performance across models or tasks.

The video also discusses research uses, including plotting residual vectors to study refusal behavior and testing how alignment training withstands modification. It closes with a warning about misuse. Hardware requirements, platform support and setup commands are not covered.