Research map / Vision-language and multimodal / Multimodal LLMs
ToolVQA: A Dataset for Multi-step Reasoning VQA with External Tools
Shaofeng Yin, Ting Lei, Yang Liu
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 LCS-based example matching (ablation) | ToolEngine · ToolEngine data quality | answer accuracy: 41.6 -> 90.8 | Improved |
| add real-world examples on top of LCS (ablation) | ToolEngine · ToolEngine data quality | answer accuracy: 27.3 -> 41.6 | Improved |
| add LCS matching (ablation) | ToolEngine · ToolEngine data quality | reasoning complexity: 1.61 -> 2.38 | Improved |
| add LCS matching (ablation) | ToolEngine · ToolEngine data quality | correlation: 81.4 -> 85.2 | No clear effect |
| add ImageCaption tool (ablation) | Tuned LLaVA-7B · ToolVQA test | accuracy: 11.1 -> 18.8 | Improved |
| add tool trajectories to training (ablation) | Tuned LLaVA-7B · ToolVQA test | accuracy: 8.01 -> 18.8 | Improved |
| add image input (ablation) | Tuned LLaVA-7B · ToolVQA test | accuracy: 0.18 -> 8.01 | Improved |
| fine-tune LLaVA-7B on ToolVQA | LLaVA-v1.5-7B · ToolVQA test (VLM + tool) | accuracy: 1.17 -> 18.8 | Improved |
| add tools to VLM for untuned models | GPT-4o · ToolVQA test | accuracy: 38.29 -> 34.96 | Hurt |
| fine-tune LLaVA-7B on ToolVQA | LLaVA-7B · OOD benchmarks | accuracy: 41.2 -> 47 | Improved |
| fine-tune LLaVA-7B on ToolVQA | LLaVA-7B · TEMPLAMA | accuracy: 3.06 -> 21.43 | Improved |
| add more in-context shots | Tuned LLaVA-7B · ToolVQA few-shot | accuracy: 18.80 -> 21.13 | Improved |
Closest papers by meaning
- WebQA: Multihop and Multimodal QA (CVPR 2022 · similarity 0.91)
- ReAG: Reasoning-Augmented Generation for Knowledge-based Visual Question Answering (CVPR 2026 · similarity 0.91)
- SK-VQA: Synthetic Knowledge Generation at Scale for Training Context-Augmented Multimodal LLMs (ICML 2025 · similarity 0.89)
- VTQA: Visual Text Question Answering via Entity Alignment and Cross-Media Reasoning (CVPR 2024 · similarity 0.89)
- When Open-Vocabulary Visual Question Answering Meets Causal Adapter: Benchmark and Approach (AAAI 2025 · similarity 0.89)
- MMFineReason: Closing the Multimodal Reasoning Gap via Open Data-Centric Methods (NeurIPS 2026 · 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.
