Research map / Large language models / Alignment and preferences
Multi-dimensional Preference Alignment by Conditioning Reward Itself
JIHO JANG, Jin-Young Kim, Kyungjune Baek, Nojun Kwak
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 dimensional reward dropout (vs rate 0.0) (ablation) | MCDPO · Pick-a-Pic v2 test (SD1.5) | CLIP win rate: 49.2 -> 57.6 | Improved |
| add dimensional reward dropout (vs rate 0.0) (ablation) | MCDPO · Pick-a-Pic v2 test (SD1.5) | Avg win rate: 80.1 -> 81.1 | Improved |
| add MCSFT pre-training (ablation) | MCDPO · Pick-a-Pic v2 test (SD1.5) | Avg win rate: 60.5 -> 81.1 | Improved |
| add context-aware reward conditioning module (ablation) | MCDPO · Pick-a-Pic v2 test (SD1.5) | Avg win rate: 70.1 -> 81.1 | Improved |
| use MCDPO alignment after MCSFT instead of full MCSFT training alone | MCSFT · Pick-a-Pic v2 test (SD1.5) | Avg win rate: 73.2 -> 81.1 | Improved |
| train MCDPO on conflict-only pairs instead of random pairs | MCDPO · Pick-a-Pic v2 test (SD1.5) | Avg win rate: 81.1 -> 77.0 | Hurt |
| train standard DPO on conflict-only pairs instead of random pairs | DPO · Pick-a-Pic v2 test (SD1.5) | Avg win rate: 66.7 -> 59.0 | Hurt |
| use MCDPO on random pairs instead of filtered conflict-free DPO | Filtered DPO · Pick-a-Pic v2 test (SD1.5) | Avg win rate: 78.7 -> 81.1 | Improved |
| use MCDPO instead of standard DPO | DPO · Pick-a-Pic v2 test (SD1.5) | Avg win rate: 66.7 -> 81.1 | Improved |
| use conditional multi-dimension MCDPO instead of Aesthetic-only Conditional DPO specialist | Cond. DPO specialist · Pick-a-Pic v2 test (SD1.5) | Aesthetic win rate: 88.1 -> 91.8 | Improved |
| amplify CLIP only at test time | MCDPO · Pick-a-Pic v2 test (SD1.5) | CLIP win rate: 57.6 -> 59.0 | Improved |
| add Aes-CLIP temporal scheduling at test time | MCDPO · Pick-a-Pic v2 test (SD1.5) | CLIP win rate: 57.6 -> 59.6 | Improved |
Closest papers by meaning
- AlphaPO: Reward Shape Matters for LLM Alignment (ICML 2025 · similarity 0.89)
- Sequential Preference Optimization: Multi-Dimensional Preference Alignment with Implicit Reward Modeling (AAAI 2025 · similarity 0.89)
- Data with High and Consistent Preference Difference Are Better for Reward Model (AAAI 2025 · similarity 0.89)
- MPO: An Efficient Post-Processing Framework for Mixing Diverse Preference Alignment (ICML 2025 · similarity 0.89)
- Rewards-in-Context: Multi-objective Alignment of Foundation Models with Dynamic Preference Adjustment (ICML 2024 · similarity 0.89)
- Reward-Augmented Data Enhances Direct Preference Alignment of LLMs (ICML 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.
