Adding Random Noise During Training Improves Alignment Generalization
Sauers_ · x · 2026-11-15
A study shows that injecting random noise during training—while instructing the model to keep its values stable under perturbation—improves alignment generalization, offering a simple approach to making safety training more robust.
Related event: Noise injection during training may improve alignment generalization(4 posts)→
More from Safety
- AI-native companies reportedly churning from Fireworks to Baseten after security incident — AccBalanced · 2026-11-15
- Plenty of AI company talent would take big pay cuts to join public safety bodies — AndyMasley · 2026-11-15
- New York's RAISE Act takes effect: frontier labs must register by Jan 1, 2027, report incidents in 72h — dl_weekly · 2026-11-15
- DC 'March for an AI Treaty' calls for banning recursive self-improvement — jeremiecharris · 2026-11-15
- OpenAI adds a setting to auto-solve CAPTCHAs forever, drawing Solana UI jokes — mertdumenci · 2026-11-15
- AI-written story laundered through news 'sources': researchers slam AP, Guardian coverage — Afinetheorem · 2026-11-15