Agent: Cursor, Claude CodeLLM: Gemma 3, LLaMA#llm#abliteration#transformer#uncensoring#model-editing
Heretic removes safety alignment from LLMs without post-training by combining directional ablation with a TPE-based optimizer (Optuna). It automatically finds optimal abliteration parameters by co-minimizing refusals and KL divergence, preserving model intelligence while eliminating censorship.