Research map / Vision-language and multimodal / Visual question answering
A Progressive Evidence Localization Framework Based on Wasserstein Gradient Flows for Document Visual Question Answering
Haosen Wang, Jing Xiao, Mengqiao Li, Xuanze Wang, Mingzheng Zhang, Xiaowang Zhang, Zhiyong Feng
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 |
|---|---|---|---|
| use multi-step Wasserstein evolution (T=6) instead of one-step localization (T=1) (ablation) | Wasserstein gradient flow localizer · DocVQA | mAP@IoU: 40.2 -> 79.6 | Improved |
| increase evolution steps from 6 to 9 (ablation) | Wasserstein gradient flow localizer · DocVQA | mAP@IoU: 79.6 -> 78.4 | Hurt |
| increase evolution steps from 2 to 4 (ablation) | Wasserstein gradient flow localizer · DocVQA | mAP@IoU: 52.8 -> 73.1 | Improved |
| raise entropy weight lambda from 1.0 to 2.0 (ablation) | Wasserstein gradient flow localizer · DocVQA | mAP@IoU: 79.6 -> 66.8 | Hurt |
| lower entropy weight lambda from 1.0 to 0.5 (ablation) | Wasserstein gradient flow localizer · DocVQA | mAP@IoU: 79.6 -> 71.4 | Hurt |
| lower time step tau from 0.1 to 0.05 (ablation) | Wasserstein gradient flow localizer · DocVQA | mAP@IoU: 79.6 -> 74.8 | Hurt |
| use progressive Wasserstein evidence localization instead of DocVXQA | DocVXQA · NiM-Benchmark | mAP@IoU: 46.5 -> 53.4 | Improved |
| use progressive localization instead of DocVXQA | DocVXQA · DocVQA | mAP@IoU: 72.1 -> 79.6 | Improved |
| use progressive localization at strict IoU 0.9 | DocVXQA · DUDE | AP@IoU 0.9: 23.4 -> 25.3 | Improved |
Closest papers by meaning
- DocVAL: Validated Chain-of-Thought Distillation for Grounded Document VQA (ICML 2026 · similarity 0.87)
- DocSeeker: Structured Visual Reasoning with Evidence Grounding for Long Document Understanding (CVPR 2026 · similarity 0.86)
- Look as You Think: Unifying Reasoning and Visual Evidence Attribution for Verifiable Document RAG via Reinforcement Learning (AAAI 2026 · similarity 0.85)
- DocVXQA: Context-Aware Visual Explanations for Document Question Answering (ICML 2025 · similarity 0.85)
- VITED: Video Temporal Evidence Distillation (CVPR 2025 · similarity 0.85)
- SIEVES: Selective Prediction Generalizes through Visual Evidence Scoring (NeurIPS 2026 · similarity 0.84)
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.
