Research map / Vision-language and multimodal / Visual question answering
Wiki-R1: Incentivizing Multimodal Reasoning for Knowledge-based VQA via Data and Sampling Curriculum
Shan Ning, Longtian Qiu, Xuming He
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 DAPO RL instead of SFT (ablation) | Qwen-2.5-VL 3B · Encyclopedic VQA / InfoSeek | overall accuracy: EVQA 25.1 -> 31.4; InfoSeek 29.5 -> 41.5 | Improved |
| Add data curriculum generation to DAPO (ablation) | DAPO on Qwen-2.5-VL 3B · Encyclopedic VQA / InfoSeek | overall accuracy: EVQA 31.4 -> 34.5; InfoSeek 41.5 -> 43.0 | Improved |
| Add curriculum sampling without observation propagation (ablation) | DAPO + data curriculum · Encyclopedic VQA / InfoSeek | overall accuracy: EVQA 34.5 -> 32.1; InfoSeek 43.0 -> 40.0 | Hurt |
| Add observation propagation to curriculum sampling (ablation) | DAPO + data + sampling curriculum · Encyclopedic VQA / InfoSeek | overall accuracy: EVQA 32.1 -> 35.9; InfoSeek 40.0 -> 42.2 | Improved |
| Full Wiki-R1 vs data curriculum alone on InfoSeek (ablation) | DAPO on Qwen-2.5-VL 3B · InfoSeek overall | accuracy: 43.0 -> 42.2 (EVQA 34.5 -> 35.9) | Mixed |
| Use Sentence Transformer instead of TF-IDF for label-propagation similarity (ablation) | Wiki-R1 3B · Encyclopedic VQA / InfoSeek | overall accuracy: EVQA 35.9 -> 36.3; InfoSeek 42.2 -> 43.0 | Improved |
| Use Wiki-R1 7B instead of best ReflectiVA retrieval mode | KB-VQA · Encyclopedic VQA (all) and InfoSeek (all) | accuracy: EVQA 35.5 -> 37.1; InfoSeek 40.1 -> 44.1 | Improved |
| Provide oracle Wikipedia entity: Wiki-R1 7B vs ReflectiVA on EVQA single-hop | Wiki-R1 7B · EVQA single-hop (oracle entity) | accuracy: 75.2 -> 69.2 | Hurt |
Closest papers by meaning
- Knowledge-based Visual Question Answer with Multimodal Processing, Retrieval and Filtering (NeurIPS 2025 · similarity 0.91)
- Augmenting Multimodal LLMs with Self-Reflective Tokens for Knowledge-based Visual Question Answering (CVPR 2025 · similarity 0.90)
- SK-VQA: Synthetic Knowledge Generation at Scale for Training Context-Augmented Multimodal LLMs (ICML 2025 · similarity 0.90)
- Separation of Powers: On Segregating Knowledge from Observation in LLM-enabled Knowledge-based Visual Question Answering (CVPR 2025 · similarity 0.90)
- Latent Visual Reasoning (ICLR 2026 · similarity 0.89)
- BOK-VQA: Bilingual outside Knowledge-Based Visual Question Answering via Graph Representation Pretraining (AAAI 2024 · similarity 0.89)
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.
