Training-Free Weight Editing Cuts Harmful Compliance: Honesty Rate 31% to 88%
AI-Interpretability-illinois · hf · 2026-10-19
A paper from an Illinois interpretability group introduces a training-free weight-editing method: using just 25 minimal prompt pairs (fictional-scenario wrong answers vs truthful answers), it contrasts input-channel norms, selects the top 4% channels per matrix, and scales those weights with α=-1.25 (about 0.8% of weights meaningfully edited). On a held-out 'ordered to lie' test, honest-response rate rises from 31% unedited, 67% with honest prompts, 83% with steering, to 88% — no retraining needed. The recipe is as simple as a few prompt pairs plus one weight scaling.
More from Safety
- Musk warns many will be 'one-shotted' by super-persuasive superintelligences — elonmusk · 2026-10-19
- Why safety reports can't prove safety: the adversarial logic flaw — davidmanheim · 2026-10-19
- JPMorgan, Citi, BofA block AI agents from accounts — approval gates are the architecture — ur_piyo_a_hoe · 2026-10-19
- Researcher: AI safety evals repeat crypto's classic 'I couldn't crack it' fallacy — davidmanheim · 2026-10-19
- Gemini's critical summary of a safety report surfaces downplayed residual risks — davidmanheim · 2026-10-19
- Red teamers share AI safety findings that seemed harmless until they weren't — Plastic_Secret1303 · 2026-10-19