Research map / Large language models / Alignment and preferences
When Human Preferences Flip: An Instance-Dependent Robust Loss for RLHF
Yifan Xu, Xichen Ye, Yifan Chen, Qiaosheng Zhang
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 warmup of the flipping model (20/20 batches) (ablation) | FA-DPO · HH_Golden 20% flip | accuracy: 84.96 -> 98.56 | Improved |
| add warmup (20/20 batches) (ablation) | FA-DPO · HH_Golden 20% flip | win rate: 57.35 -> 70.04 | Improved |
| add warmup (50/50 batches) (ablation) | FA-DPO · HH_Golden 20% flip | accuracy: 83.16 -> 96.80 | Improved |
| more policy steps (20 -> 50) with warmup (ablation) | FA-DPO · HH_Golden 20% flip | win rate: 70.04 -> 82.90 | Improved |
| more policy steps (20 -> 50) without warmup (ablation) | FA-DPO · HH_Golden 20% flip | accuracy: 84.96 -> 77.00 | Hurt |
| more policy steps (20 -> 50) with warmup (ablation) | FA-DPO · HH_Golden 20% flip | accuracy: 98.56 -> 98.40 | No clear effect |
| replace DPO with FA-DPO | Pythia-1B · UltraFeedback 40% flip | accuracy: 51.87 -> 70.77 | Improved |
| replace DPO with FA-DPO | Pythia-1B · HH_Golden 40% flip | accuracy: 58.63 -> 98.83 | Improved |
| replace DPO with FA-DPO | Pythia-1B · UltraFeedback clean (0%) | win rate: 67.00 -> 66.85 | Hurt |
| replace DPO with FA-DPO | Llama-3.1-8B · UltraFeedback 40% flip | accuracy: 64.96 -> 78.87 | Improved |
| replace DPO with FA-DPO | Mistral-7B · UltraFeedback 40% flip | win rate: 35.35 -> 59.00 | Improved |
| replace ROPO with FA-DPO (clean data) | Pythia-1B · UltraFeedback clean (0%) | accuracy: 70.75 -> 73.05 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- ROPO — ROPO: Robust Preference Optimization for Large Language Models (ICML 2025) · better on 5 of 5 shared results
- SIMPO — SimPO: Simple Preference Optimization with a Reference-Free Reward (NeurIPS 2024) · better on 1 of 1 shared result
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
- Online Preference Alignment for Language Models via Count-based Exploration (ICLR 2025 · similarity 0.92)
- Iterative Preference Learning from Human Feedback: Bridging Theory and Practice for RLHF under KL-constraint (ICML 2024 · similarity 0.92)
- Doubly Robust Alignment for Large Language Models (NeurIPS 2025 · similarity 0.91)
- Beyond Reverse KL: Generalizing Direct Preference Optimization with Diverse Divergence Constraints (ICLR 2024 · similarity 0.91)
- Linear Alignment: A Closed-form Solution for Aligning Human Preferences without Tuning and Feedback (ICML 2024 · similarity 0.91)
- Data with High and Consistent Preference Difference Are Better for Reward Model (AAAI 2025 · similarity 0.91)
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.
