Research map / Large language models / Alignment and preferences
Robust Preference Optimization: Aligning Language Models with Noisy Preference Feedback
Xiaoyang Cao, Zelai Xu, Mo Guang, Kaiwen Long, Michiel A. Bakker, Yu Wang, Chao Yu
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 |
|---|---|---|---|
| add RE-PO reliability reweighting to DPO | DPO on Mistral-7B-Instruct-v0.2 · AlpacaEval 2 (UltraFeedback-trained) | LC / WR (%): 28.5/28.6 -> 35.5/33.0 | Improved |
| add RE-PO to IPO | IPO on Llama-3-8B-Instruct · AlpacaEval 2 (UltraFeedback-trained) | LC / WR (%): 43.6/41.6 -> 48.3/48.6 | Improved |
| add RE-PO to DPO | DPO on Llama-3-8B-Instruct · AlpacaEval 2 (UltraFeedback-trained) | LC / WR (%): 40.8/42.9 -> 44.1/46.2 | Improved |
| add RE-PO to SimPO | SimPO on Llama-3-8B-Instruct · AlpacaEval 2 (UltraFeedback-trained) | LC / WR (%): 44.5/37.1 -> 46.9/39.4 | Mixed |
| add RE-PO to CPO | CPO on Mistral-7B-Instruct-v0.2 · AlpacaEval 2 (UltraFeedback-trained) | LC / WR (%): 26.3/26.4 -> 27.6/27.8 | No clear effect |
| add label smoothing to SimPO | SimPO on Llama-3-8B-Instruct · AlpacaEval 2 (UltraFeedback-trained) | LC (%): 44.5 -> 48.1 | Improved |
| add label smoothing to IPO | IPO on Llama-3-8B-Instruct · AlpacaEval 2 (UltraFeedback-trained) | LC / WR (%): 43.6/41.6 -> 40.3/38.2 | Hurt |
| use RE-DPO instead of rDPO and Holder-DPO | DPO on Llama-3-8B-Instruct · AlpacaEval 2 (UltraFeedback-trained) | LC / WR (%): 37.3/35.4 -> 44.1/46.2 | Improved |
| add RE-DPO with per-annotator reliabilities on MultiPref | DPO on Llama-3-8B-Instruct · AlpacaEval 2 (MultiPref-trained) | LC / WR (%): 36.7/39.3 -> 41.1/44.4 | Improved |
| add RE-DPO with per-annotator reliabilities on MultiPref | DPO on Mistral-7B-Instruct-v0.2 · AlpacaEval 2 (MultiPref-trained) | LC / WR (%): 28.8/26.4 -> 31.8/28.8 | Improved |
| set initial reliability eta0 to 0.9 instead of 0.99 (ablation) | RE-DPO Mistral-7B · AlpacaEval 2 (UltraFeedback-trained) | LC (%): 30.9 -> 35.5 | Improved |
| set initial reliability eta0 to 0.9 instead of 0.55 (ablation) | RE-DPO Mistral-7B · AlpacaEval 2 (UltraFeedback-trained) | LC (%): 31.4 -> 35.5 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SimPO — SimPO: Simple Preference Optimization with a Reference-Free Reward (NeurIPS 2024) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Provably Robust DPO: Aligning Language Models with Noisy Feedback (ICML 2024 · similarity 0.94)
- Towards Robust Alignment of Language Models: Distributionally Robustifying Direct Preference Optimization (ICLR 2025 · similarity 0.94)
- Doubly Robust Alignment for Large Language Models (NeurIPS 2025 · similarity 0.94)
- ROPO: Robust Preference Optimization for Large Language Models (ICML 2025 · similarity 0.93)
- ComPO: Preference Alignment via Comparison Oracles (NeurIPS 2025 · similarity 0.93)
- Towards Efficient Exact Optimization of Language Model Alignment (ICML 2024 · similarity 0.93)
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.
