Heretic hits 42k stars: training-free weight editing kills refusals in 30 minutes
aigclink · x · 2026-11-19
Heretic, an open-source GitHub project with 42k stars, modifies local LLM behavior with no fine-tuning or retraining.
How it works
- Classic path Abliteration: identify refusal-related activation directions and edit weights directly; v2.0 defaults to the more flexible ARA, dropping the single-refusal-direction assumption
- Optuna auto-searches modification parameters to change target behavior while preserving other capabilities
Results
- A 4B Qwen processes in 30–40 minutes on an RTX 3090
- Gemma 3 12B refusal rate drops from 97% to 3%
Beyond de-censoring and removing slop/sycophancy, it ships ready-made configs (less humor, different emotional alignment, terser replies), AGPL-3.0 licensed.
More from Models
- Leaked Gemini 4 Neon benchmarks show sweeps across 17 evals vs Claude, GPT — nevrekaraishwa2 · 2026-11-19
- Kimi K3.1 Pro beats K3 on 14 of 17 benchmarks; pairing with Opus 5.5 halves spend — Kimi_Moonshot · 2026-11-19
- Simular box hits OSWorld v2 SOTA at 80.7% success, $1.79 per task — 8x cheaper than August — xwang_lk · 2026-11-19
- Peking University Releases Foundation Model for Measuring and Predicting Human Behaviour — PekingUniversity · 2026-11-19
- Heretic, a 42k-star open-source tool, cuts Gemma 3 12B refusal rate from 97% to 3% — aigclink · 2026-11-19
- Jeremy Howard reverts Solveit models to pre-6.2 versions over quality issues — jeremyphoward · 2026-11-19