A Unified Theoretical Analysis of Private and Robust Offline Alignment: from RLHF to DPO

Documenti analoghi

• Offline and Online KL-Regularized RLHF under Differential Privacy
  di: Wu, Yulian, et al.
  Pubblicazione: (2025)
• Square$χ$PO: Differentially Private and Robust $χ^2$-Preference Optimization in Offline Direct Alignment
  di: Zhou, Xingyu, et al.
  Pubblicazione: (2025)
• Provably Mitigating Corruption, Overoptimization, and Verbosity Simultaneously in Offline and Online RLHF/DPO Alignment
  di: Chen, Ziyi, et al.
  Pubblicazione: (2025)
• Improved Bounds for Private and Robust Alignment
  di: Weng, Wenqian, et al.
  Pubblicazione: (2025)
• Conditional Equivalence of DPO and RLHF: Implicit Assumption, Failure Modes, and Provable Alignment
  di: Yang, Zhiqin, et al.
  Pubblicazione: (2026)