Heretic 2.0 Released: General-Purpose LLM Modification Framework
Heretic 2.0 has been released as a major update, evolving from a censorship-removal tool into a general-purpose LLM modification framework. It introduces Arbitrary-Rank Ablation (ARA) to replace traditional directional abliteration and adopts a modular architecture built around Python plugins. This release significantly broadens the capabilities of open-source model tuning by enabling developers to suppress slop, reduce sycophancy, and alter emotional alignment alongside standard censorship removal. It empowers the community to fine-tune model behaviors effectively without heavy retraining. Heretic 2.0 introduces Arbitrary-Rank Ablation (ARA) which uses the L-BFGS optimizer to steer model representations without rank restrictions. The software has been completely restructured around customizable scorer and modifier plugins, though it is not backwards compatible with Heretic 1.x configurations.
## BACKGROUND
Large language models are typically aligned using techniques like Reinforcement Learning from Human Feedback (RLHF), which can introduce rigid refusal behaviors or unwanted conversational traits such as sycophancy. Mechanistic interpretability research previously showed that safety refusals are often mediated by specific linear directions in the model's activation space, a concept that tools like Heretic modify to alter model behavior.