Research map / Vision-language and multimodal / Multimodal LLMs
Direct Preference Optimization for Perceptual Alignment via Vision-Language Consistency
Seungyeon Lee, Donggyu Lee
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 description-based responses (ablation) | VLC-DPO (LLaVA-1.5 7B) · AMBER (Disc.) / POPE | F1 / accuracy: 82.0 -> 87.2 F1; 82.6 -> 85.2 POPE | Improved |
| add consistency filtering of response pairs (ablation) | VLC-DPO (LLaVA-1.5 7B) · AMBER (Disc.) / POPE | F1 / accuracy: 81.7 -> 87.2 F1; 81.8 -> 85.2 POPE | Improved |
| add VLC-score difference penalty weight |wc| (ablation) | VLC-DPO (LLaVA-1.5 7B) · AMBER (Disc.) / POPE | F1 / accuracy: 77.9 -> 87.2 F1; 81.0 -> 85.2 POPE | Improved |
| raise consistency threshold from 0.5 to 0.7 (ablation) | VLC-DPO (LLaVA-1.5 7B) · AMBER (Disc.) / POPE | F1 / accuracy: 82.9 -> 87.2 F1; 83.2 -> 85.2 POPE | Improved |
| raise consistency threshold from 0.7 to 0.8 (ablation) | VLC-DPO (LLaVA-1.5 7B) · AMBER (Disc.) / POPE | F1 / accuracy: 87.2 -> 86.1 F1; 85.2 -> 84.7 POPE | Hurt |
| add mixture noise injection to rejected responses (vs noise-free) (ablation) | VLC-DPO (LLaVA-1.5 7B) · AMBER (Disc.) / POPE | F1 / accuracy: 81.4 -> 87.2 F1; 81.6 -> 85.2 POPE | Improved |
| use order-shuffling noise only instead of noise-free (ablation) | VLC-DPO (LLaVA-1.5 7B) · AMBER (Disc.) / POPE | F1 / accuracy: 81.4 -> 79.0 F1; 81.6 -> 81.3 POPE | Hurt |
| use replacement noise only instead of mixture (ablation) | VLC-DPO (LLaVA-1.5 7B) · AMBER (Disc.) / POPE | F1 / accuracy: 87.2 -> 86.1 F1; 85.2 -> 83.9 POPE | Hurt |
| apply VLC-DPO to LLaVA-1.5 7B | VLC-DPO (LLaVA-1.5 7B) · POPE (Adversarial) | accuracy: 80.8 -> 85.2 | Improved |
| apply VLC-DPO to LLaVA-1.5 13B | VLC-DPO (LLaVA-1.5 7B) · A-OKVQA (zero-shot) | VQA score: 52.3 -> 66.5 | Improved |
| apply VLC-DPO to LLaVA-1.5 7B | VLC-DPO (LLaVA-1.5 7B) · OK-VQA (zero-shot) | VQA score: 46.4 -> 57.3 | Improved |
| apply VLC-DPO to LLaVA-1.5 7B | VLC-DPO (LLaVA-1.5 7B) · MMHal-Bench | score: 1.86 -> 2.95 | Improved |
Closest papers by meaning
- MIA-DPO: Multi-Image Augmented Direct Preference Optimization For Large Vision-Language Models (ICLR 2025 · similarity 0.94)
- Fine-Grained Verifiers: Preference Modeling as Next-token Prediction in Vision-Language Alignment (ICLR 2025 · similarity 0.92)
- RoVRM: A Robust Visual Reward Model Optimized via Auxiliary Textual Preference Data (AAAI 2025 · similarity 0.91)
- Calibrated Self-Rewarding Vision Language Models (NeurIPS 2024 · similarity 0.90)
- Multimodal Large Language Models Make Text-to-Image Generative Models Align Better (NeurIPS 2024 · similarity 0.90)
- Task Preference Optimization: Improving Multimodal Large Language Models with Vision Task Alignment (CVPR 2025 · similarity 0.90)
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.
