Robust Optimization for Mitigating Reward Hacking with Correlated Proxies

Documenti analoghi

• Enhancing LLM Safety via Constrained Direct Preference Optimization
  di: Liu, Zixuan, et al.
  Pubblicazione: (2024)
• Correlated Proxies: A New Definition and Improved Mitigation for Reward Hacking
  di: Laidlaw, Cassidy, et al.
  Pubblicazione: (2024)
• Belief-Enriched Pessimistic Q-Learning against Adversarial State Perturbations
  di: Sun, Xiaolin, et al.
  Pubblicazione: (2024)
• Reward Hacking Mitigation using Verifiable Composite Rewards
  di: Tarek, Mirza Farhan Bin, et al.
  Pubblicazione: (2025)
• Repairing Reward Functions with Feedback to Mitigate Reward Hacking
  di: Hatgis-Kessell, Stephane, et al.
  Pubblicazione: (2025)