Research map / Large language models / Alignment and preferences
Iterative Perceptual Alignment for VLMs via Deterministic Reconstruction Feedback
Xiaorui Chen, Hanzhong Guo, Nizhe Cai, Jieliang Luo
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 |
|---|---|---|---|
| DPO on reconstruction-feedback preference data (>0.03, 5K) | Qwen3-VL-2B-Instruct · Hallusion | accuracy (%): 46.27 -> 57.83 | Improved |
| DPO on reconstruction-feedback data | InternVL3.5-2B-HF · POPE | accuracy (%): 86.67 -> 88.66 | Improved |
| DPO on reconstruction-feedback data | InternVL3.5-1B-HF · Hallusion | accuracy (%): 54.47 -> 51.10 | Hurt |
| DPO on reconstruction-feedback data | InternVL3.5-1B-HF · MM-Bench | accuracy (0-1): 0.32 -> 0.49 | Improved |
| DPO on reconstruction-feedback data | InternVL3.5-2B-HF · RealWorldQA | accuracy (0-1): 0.39 -> 0.35 | Hurt |
| use 3K auto data (>0.05) instead of 5K (>0.03) (ablation) | DPO on Qwen3-VL-2B · MM-Bench | accuracy (0-1): 0.67 -> 0.70 | Improved |
| use 5K auto data (>0.03) instead of 3K (>0.05) (ablation) | DPO on Qwen3-VL-2B · Hallusion | accuracy (%): 56.99 -> 57.83 | Improved |
| replace human-annotated RLHF-V 3.5k with automatic 3K data | DPO on InternVL3.5-2B · POPE | accuracy (%): 88.56 -> 88.66 | No clear effect |
| train reward model on reconstruction-based pairs (about 20% of data) | T2I-Consistency-Score · zero-shot RLHF-V / VLFeedback | accuracy (%): 74.12 vs 65.25 (RLHF-V); 67.43 vs 77.41 (POVID) | Mixed |
| add score decomposition with text-only branch | reward model · in-domain preference | accuracy | No clear effect |
Closest papers by meaning
- Fine-Grained Verifiers: Preference Modeling as Next-token Prediction in Vision-Language Alignment (ICLR 2025 · similarity 0.90)
- Deep Pre-Alignment for VLMs (ICML 2026 · similarity 0.88)
- Gated Relational Alignment via Confidence-based Distillation for Efficient VLMs (ICML 2026 · similarity 0.88)
- LVLM-Aided Alignment of Task-Specific Vision Models (CVPR 2026 · similarity 0.88)
- Towards Self-Refinement of Vision-Language Models with Triangular Consistency (NeurIPS 2025 · similarity 0.87)
- Amend to Alignment: Decoupled Prompt Tuning for Mitigating Spurious Correlation in Vision-Language Models (ICML 2024 · similarity 0.87)
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.
