Safety Training Modulates Harmful Misalignment Under On-Policy RL, But Direction Depends on Environment Design

Documenti analoghi

• Short-circuiting Shortcuts: Mechanistic Investigation of Shortcuts in Text Classification
  di: Eshuijs, Leon, et al.
  Pubblicazione: (2025)
• Balancing the Scales: Reinforcement Learning for Fair Classification
  di: Eshuijs, Leon, et al.
  Pubblicazione: (2024)
• Efficient Jailbreaking of Large Models by Freeze Training: Lower Layers Exhibit Greater Sensitivity to Harmful Content
  di: Shen, Hongyuan, et al.
  Pubblicazione: (2025)
• Investigating the Effect of Misalignment on Membership Privacy in the White-box Setting
  di: Cretu, Ana-Maria, et al.
  Pubblicazione: (2023)
• The Surprising Harmfulness of Benign Overfitting for Adversarial Robustness
  di: Hao, Yifan, et al.
  Pubblicazione: (2024)