Research map / Large language models / Alignment and preferences
Preference Robustness for DPO with Applications to Public Health
Cheol Woo Kim, Shresth Verma, Mauricio Tec, Milind Tambe
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use DPO-PRO (rho=0.03) instead of DPO under high label-flip noise | DPO · UltraFeedback, high noise (alpha=0.6) | win rate (%): 26.97 -> 28.98 | Improved |
| use DPO-PRO (rho=0.008) instead of DPO with no noise | DPO · UltraFeedback, no noise | win rate (%): 43.77 -> 43.24 | Hurt |
| use DPO-PRO (rho=0.03) instead of DrDPO under low noise | DrDPO · UltraFeedback, low noise (alpha=0.3) | win rate (%): 38.84 -> 33.33 | Hurt |
| use DPO-PRO (rho=0.03) instead of DPO on evaluation reward under high noise | DPO · UltraFeedback, high noise (alpha=0.6) | evaluation reward: 400.44 -> 499.59 | Improved |
| use DPO-PRO (rho=0.1) instead of DPO on evaluation reward with no noise | DPO · UltraFeedback, no noise | evaluation reward: 877.01 -> 781.28 | Hurt |
| use DPO-PRO (rho=0.008) instead of DPO on evaluation reward under low noise | DPO · UltraFeedback, low noise (alpha=0.3) | evaluation reward: 729.12 -> 713.00 | Hurt |
| increase robustness radius rho (ablation) | DPO-PRO · UltraFeedback, win rate | win rate (%): rho 0.008: 43.24 -> 28.75; rho 0.03: 41.71 -> 28.98 | Mixed |
| use DPO-PRO (rho=0.3) instead of DPO on ground-truth reward, low noise | DPO · ARMMAN RMAB, low noise | avg ground-truth reward: 5707.09 -> 5904.7 | Improved |
| use DPO-PRO (rho=0.3) instead of DPO on ground-truth reward, no noise | DPO · ARMMAN RMAB, no noise | avg ground-truth reward: 5465.36 -> 5797.1 | Improved |
| use DPO-PRO (rho=0.03) instead of DPO on win rate, low noise | DPO · ARMMAN RMAB, low noise | win rate (%): 13.84 -> 13.5 | No clear effect |
| use DPO-PRO (rho=0.3) instead of DPO on win rate, low noise | DPO · ARMMAN RMAB, low noise | win rate (%): 13.84 -> 17.8 | Improved |
| use DPO-PRO (rho=0.1) instead of DPO on win rate, no noise | DPO · ARMMAN RMAB, no noise | win rate (%): 15.38 -> 17.7 | Improved |
Closest papers by meaning
- Towards Robust Alignment of Language Models: Distributionally Robustifying Direct Preference Optimization (ICLR 2025 · similarity 0.88)
- What Matters in Data for DPO? (NeurIPS 2025 · similarity 0.88)
- Group Robust Preference Optimization in Reward-free RLHF (NeurIPS 2024 · similarity 0.88)
- Provably Robust DPO: Aligning Language Models with Noisy Feedback (ICML 2024 · similarity 0.88)
- MaPPO: Maximum a Posteriori Preference Optimization with Prior Knowledge (NeurIPS 2026 · similarity 0.88)
- 3D-Properties: Identifying Challenges in DPO and Charting a Path Forward (ICLR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
