AI BriefingHeretic 2.0 Released: General-Purpose Model Modification Tool with Arbitrary-Rank Ablation
·2026.11.01 22:16
Key point
The update replaces directional abliteration with Arbitrary-Rank Ablation (ARA) and introduces a plugin system for custom scoring and steering.
Details
- Heretic 2.0 transforms the tool from a censorship removal utility into a general-purpose model modification platform, capable of suppressing slop, reducing sycophancy, and altering emotional alignment.
- The new default steering method, Arbitrary-Rank Ablation (ARA), uses the L-BFGS optimizer to modify matrices without rank restrictions, outperforming the previous directional abliteration approach.
- A new plugin architecture allows users to define custom scorer plugins for optimization objectives, including a BERT classifier for more accurate refusal detection than keyword matching.
- The project reports 37k GitHub stars, over 8,500 models on Hugging Face, and 88 million total model downloads, with improved support for thinking models and hyper-connections.
- Heretic 2.0 is not backwards compatible with 1.x configurations, requiring users to follow a migration guide to update their setups.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Prospero: A lightweight C++ coding harness for local LLM agentsspinf-serve: Open-source inference server for bulk LLM scoring on single GPU
Herdr releases native desktop app for managing terminals and coding agents