Today’s brief · 2026-09-01
Heretic review: automated abliteration that hits 3/100 refusals at 0.16 KL on a 12B model
Heretic automates abliteration: an Optuna-driven search cuts refusals to 3/100 on gemma-3-12b-it while keeping KL divergence at 0.16 — no transformer-internals knowledge required.