Research map / Vision-language and multimodal / Multimodal LLMs
Instruction-Oriented Preference Alignment for Enhancing Multi-Modal Comprehension Capability of MLLMs
Zitian Wang, Yue Liao, Kang Rong, Fengyun Rao, Yibo Yang, Si Liu
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 contrastive sampling (perturbed-image responses) (ablation) | IPA Round 1 · Ablation average (7 benchmarks) | average score: 69.1 -> 70.1 | Improved |
| add contrastive sampling (ablation) | IPA Round 1 · HallusionBench | score: +2.9 on HallusionBench | Improved |
| add Round 2 self-improvement of generator and reviser (ablation) | IPA · Ablation average (7 benchmarks) | average score: 70.1 -> 70.4 | Improved |
| add Round 3 reference-guided feedback (ablation) | IPA · Ablation average (7 benchmarks) | average score: 70.4 -> 70.5 | Improved |
| use reference-guided sampling instead of reference-guided feedback in Round 3 (ablation) | IPA · Ablation average (7 benchmarks) | average score: 70.5 -> 69.8 | Hurt |
| add full IPA pipeline to Qwen2VL-7B | Qwen2VL-7B · HallusionBench | score: 50.0 -> 54.3 | Improved |
| add full IPA pipeline to Qwen2VL-7B | Qwen2VL-7B · LLaVABench | score: 76.6 -> 84.0 | Improved |
| add full IPA pipeline to Qwen2VL-7B | Qwen2VL-7B · MMHal-Bench | score: 3.6 -> 3.7 | No clear effect |
| apply IPA preference data to Qwen2.5VL-7B | Qwen2.5VL-7B · LLaVABench | score: 75.2 -> 87.3 | Improved |
| apply IPA preference data to InternVL3-14B | InternVL3-14B · LLaVABench | score: 83.0 -> 89.9 | Improved |
| apply IPA preference data to Qwen2.5VL-72B | Qwen2.5VL-72B · MMStar | score: 71.6 -> 71.5 | No clear effect |
| align with IPA versus MMPR | Qwen2VL-7B · MMStar | score: 58.5 -> 61.7 | 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.
- RLHF-V — RLHF-V: Towards Trustworthy MLLMs via Behavior Alignment from Fine-grained Correctional Human Feedback (CVPR 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
- Automated Multi-level Preference for MLLMs (NeurIPS 2024 · similarity 0.89)
- MM-RLHF: The Next Step Forward in Multimodal LLM Alignment (ICML 2025 · similarity 0.89)
- CHiP: Cross-modal Hierarchical Direct Preference Optimization for Multimodal LLMs (ICLR 2025 · similarity 0.89)
- Mitigating Visual Hallucinations via Semantic Curriculum Preference Optimization in MLLMs (ICML 2026 · similarity 0.89)
- On-the-fly Preference Alignment via Principle-Guided Decoding (ICLR 2025 · similarity 0.88)
- ComPO: Preference Alignment via Comparison Oracles (NeurIPS 2025 · similarity 0.88)
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.
