Research map / Large language models / Alignment and preferences
UrbanAlign: Post-hoc Semantic Calibration for VLM-Human Preference Alignment
Yecheng Zhang, RONG ZHAO, Zhizhou Sha, Yong Li, Lei Wang, Ce Hou, Wen Ji, HUANG HAO, yunshan wan, Jian Yu, Junhao Xia, Yuru Zhang and 1 more
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 LWRR calibration on top of raw Mode 4 scores (ablation) | UrbanAlign · Place Pulse 2.0 (6-category avg) | accuracy (%): 55.9 -> 72.2 | Improved |
| add LWRR calibration (ablation) | UrbanAlign · Place Pulse 2.0 boring | accuracy (%): 38.9 -> 70.2 | Improved |
| add LWRR calibration (ablation) | UrbanAlign · Place Pulse 2.0 depressing | accuracy (%): 42.9 -> 68.2 | Improved |
| add multi-agent deliberation to pairwise scoring (ablation) | UrbanAlign · Place Pulse 2.0 (6-category avg) | accuracy (%): 51.9 -> 72.2 | Improved |
| add multi-agent deliberation to single-image scoring (ablation) | UrbanAlign · Place Pulse 2.0 (6-category avg) | accuracy (%): 50.9 -> 55.9 | Improved |
| add pairwise context without multi-agent deliberation (ablation) | Single-shot scoring · Place Pulse 2.0 (6-category avg) | accuracy (%): 50.9 -> 51.9 | No clear effect |
| replace zero-shot holistic prompting with structured dimension scoring (raw) (ablation) | GPT-4o · Place Pulse 2.0 (6-category avg) | accuracy (%): 56.7 -> 55.4 | No clear effect |
| add multi-agent deliberation to structured dimension scoring (raw, no LWRR) (ablation) | UrbanAlign Mode 2 · Place Pulse 2.0 (6-category avg) | accuracy (%): 55.4 -> 55.9 | No clear effect |
| replace global ridge on semantic differences with local LWRR (ablation) | UrbanAlign · Place Pulse 2.0 (6-category avg) | accuracy (%): 61.2 -> 72.2 | Improved |
| increase neighbourhood size K_max from 10 to 20 (ablation) | LWRR · Place Pulse 2.0 avg (default config) | accuracy (%): 54.1 -> 58.4 | Improved |
| raise CLIP weight alpha from 0.1 to 0.3 in hybrid space (ablation) | LWRR · Place Pulse 2.0 avg (default config) | accuracy (%): 55.9 -> 58.4 | Improved |
| increase equal threshold epsilon from 0.3 to 2.0 (ablation) | LWRR · Place Pulse 2.0 avg | accuracy (%): 60.9 -> 63.0 | Improved |
Closest papers by meaning
- LVLM-Aided Alignment of Task-Specific Vision Models (CVPR 2026 · similarity 0.87)
- UrbanFeel:A Comprehensive Benchmark for Temporal and Perceptual Understanding of City Scenes through Human Perspective (ICLR 2026 · similarity 0.87)
- WildVision: Evaluating Vision-Language Models in the Wild with Human Preferences (NeurIPS 2024 · similarity 0.87)
- SPA-VL: A Comprehensive Safety Preference Alignment Dataset for Vision Language Models (CVPR 2025 · similarity 0.86)
- Revealing the Gap in Human and VLM Scene Perception through Counterfactual Semantic Saliency (NeurIPS 2026 · similarity 0.86)
- MIA-DPO: Multi-Image Augmented Direct Preference Optimization For Large Vision-Language Models (ICLR 2025 · similarity 0.86)
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.
