Research map / Vision-language and multimodal / Visual question answering
ReasonVQA: A Multi-hop Reasoning Benchmark with Structural Knowledge for Visual Question Answering
Duong T. Tran, Trung-Kien Tran, Manfred Hauswirth, Danh Le Phuoc
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 |
|---|---|---|---|
| Fine-tune with LoRA on the ReasonVQA train set | Qwen2-VL-7B-Instruct · ReasonVQA test | accuracy: 59.0 -> 65.0 (+10.1%) | Improved |
| Fine-tune with LoRA on the ReasonVQA train set | PaliGemma-2-3B-Mix · ReasonVQA test | accuracy: 40.1 -> 66.8 (+66.5%) | Improved |
| Fine-tune with LoRA on the ReasonVQA train set | PaliGemma-2-10B-Mix · ReasonVQA test | accuracy: 65.6 -> 74.5 (+13.5%) | Improved |
| Apply answer distribution balancing (ReasonVQA-U to ReasonVQA-B) | GPT-4o · ReasonVQA open-ended | accuracy: 62.8 -> 60.8 | Hurt |
| Apply answer distribution balancing (ReasonVQA-U to ReasonVQA-B) | SmolVLM-Instruct · ReasonVQA open-ended | accuracy: 34.5 -> 31.7 | Hurt |
| Apply answer distribution balancing (ReasonVQA-U to ReasonVQA-B) | mPLUG-Owl3 · ReasonVQA open-ended | accuracy: 47.9 -> 44.8 | Hurt |
| Apply answer distribution balancing (ReasonVQA-U to ReasonVQA-B) | LLaVA-OV · ReasonVQA open-ended | accuracy: 50.5 -> 50.6 | No clear effect |
| Switch from open-ended to multiple-choice evaluation | GPT-4o · ReasonVQA-U | accuracy: 62.8 -> 76.6 | Improved |
| Switch from open-ended to multiple-choice evaluation | LLaVA-OV · ReasonVQA-U | accuracy: 50.5 -> 56.8 | Improved |
| Evaluate on ReasonVQA instead of OK-VQA | PaliGemma-2-Mix · ReasonVQA-U vs OK-VQA | accuracy: 86.8 -> 43.7 | Hurt |
| Evaluate on ReasonVQA-U instead of VQAv2 | mPLUG-Owl2 · ReasonVQA-U vs VQAv2 | accuracy: 79.4 -> 22.1 | Hurt |
Closest papers by meaning
- A-OKVQA: A Benchmark for Visual Question Answering Using World Knowledge (ECCV 2022 · similarity 0.94)
- OK-VQA: A Visual Question Answering Benchmark Requiring External Knowledge (CVPR 2019 · similarity 0.93)
- FVQA: Fact-Based Visual Question Answering (TPAMI 2017 · similarity 0.93)
- When Open-Vocabulary Visual Question Answering Meets Causal Adapter: Benchmark and Approach (AAAI 2025 · similarity 0.92)
- Separation of Powers: On Segregating Knowledge from Observation in LLM-enabled Knowledge-based Visual Question Answering (CVPR 2025 · similarity 0.92)
- VTQA: Visual Text Question Answering via Entity Alignment and Cross-Media Reasoning (CVPR 2024 · similarity 0.92)
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.
