Research map / Large language models / Alignment and preferences
Learning from Noisy Preferences: A Semi-Supervised Learning Approach to Direct Preference Optimization
Xinxin Liu, Ming Li, Zonglin Lyu, Yuzhang Shang, Chen Chen
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 first self-training iteration of pseudo-labels on noisy data after clean cold-start (ablation) | Semi-DPO (SD1.5) · HPS v2 prompts | ImageReward: 0.569 -> 0.798 | Improved |
| add second self-training iteration (ablation) | Semi-DPO (SD1.5) · HPS v2 prompts | ImageReward: 0.798 -> 0.816 | Improved |
| add second self-training iteration (ablation) | Semi-DPO (SD1.5) · HPS v2 prompts | CLIP: 0.300 -> 0.299 | No clear effect |
| add second iteration (training cost) (ablation) | Semi-DPO · training cost | GPU hours: 132 -> 228 | Hurt |
| enlarge consensus from 2 to 5 reward models (ablation) | Semi-DPO Iter0 (SD1.5) · HPS v2 prompts | ImageReward: 0.442 -> 0.569 | Improved |
| enlarge consensus from 2 to 5 reward models (ablation) | Semi-DPO Iter0 (SD1.5) · HPS v2 prompts | MPS: 12.891 -> 13.039 | Improved |
| enlarge consensus from 2 to 5 reward models (ablation) | Semi-DPO Iter0 (SD1.5) · Pick-a-Pic V2 prompts | ImageReward: 0.417 -> 0.563 | Improved |
| use dynamic per-timestep thresholds to filter pseudo-labels (vs full Diffusion-DPO) (ablation) | Semi-DPO · training cost | GPU hours: 192 -> 132 | Improved |
| replace Diff-DPO with Semi-DPO (ablation) | SD1.5 · GenEval | overall score: 43.00 -> 47.31 | Improved |
| replace Diff-DPO with Semi-DPO (ablation) | SDXL · GenEval | overall score: 58.02 -> 58.41 | Improved |
| replace Diff-DPO with Semi-DPO (ablation) | SDXL · HPS v2 prompts | PickScore: 22.953 -> 22.816 | Hurt |
| replace Diff-KTO with Semi-DPO (ablation) | SD1.5 · HPS v2 prompts | ImageReward: 0.690 -> 0.816 | Improved |
Closest papers by meaning
- Diffusion Model as a Noise-Aware Latent Reward Model for Step-Level Preference Optimization (NeurIPS 2025 · similarity 0.90)
- Smoothed Preference Optimization via ReNoise Inversion for Aligning Diffusion Models with Varied Human Preferences (ICML 2025 · similarity 0.90)
- Aesthetic Post-Training Diffusion Models from Generic Preferences with Step-by-step Preference Optimization (CVPR 2025 · similarity 0.90)
- Learning Multi-Dimensional Human Preference for Text-to-Image Generation (CVPR 2024 · similarity 0.90)
- Diffusion Model Alignment Using Direct Preference Optimization (CVPR 2024 · similarity 0.89)
- MIA-DPO: Multi-Image Augmented Direct Preference Optimization For Large Vision-Language Models (ICLR 2025 · similarity 0.89)
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.
