Research map / Vision-language and multimodal / Visual question answering
Exploring Question Decomposition for Zero-Shot VQA
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Apply decomposition only when the initial answer confidence is below a threshold | BLIP-2 13B · SLAKE | accuracy: +10.37 | Helped |
| Apply decomposition only when the initial answer confidence is below a threshold | BLIP-2 13B · PathVQA | accuracy: +22.62 | Helped |
| Apply decomposition only when the initial answer confidence is below a threshold | BLIP-2 13B · Winoground-VQA | accuracy: +22.12 | Helped |
| Apply decomposition to every question rather than selecting low-confidence answers | VQA models using generated decompositions · VQA tasks | accuracy | Hurt |
| Add human-written subquestions and their ground-truth answers | BLIP-2 3B · VQA-Introspect validation | overall accuracy: 79 -> 88.6 | Helped |
| Replace ground-truth subquestion answers with model-generated answers (ablation) | BLIP-2 3B using oracle subquestions · VQA-Introspect validation | overall accuracy: 88.6 -> 84 | Hurt |
| Remove subquestion answers from human-written decompositions (ablation) | BLIP-2 3B using oracle subquestions · VQA-Introspect validation | overall accuracy: 88.6 -> 83.3 | Hurt |
| Scramble the words in the human-written subquestion-answer context (ablation) | BLIP-2 3B · VQA-Introspect validation | overall accuracy: 88.6 -> 84.9 | Hurt |
| Use language-only rather than image-and-text decomposition generation (ablation) | Decomposers across evaluated VQA datasets · A-OKVQA, ArtVQA, OK-VQA, SLAKE | error correction rate | Mixed |
Closest papers by meaning
- Don't Just Assume; Look and Answer: Overcoming Priors for Visual Question Answering (CVPR 2018 · similarity 0.93)
- Answer Them All! Toward Universal Visual Question Answering Models (CVPR 2019 · similarity 0.93)
- Making the v in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering (CVPR 2017 · similarity 0.93)
- Debiased Visual Question Answering from Feature and Sample Perspectives (NeurIPS 2021 · similarity 0.92)
- When Open-Vocabulary Visual Question Answering Meets Causal Adapter: Benchmark and Approach (AAAI 2025 · similarity 0.92)
- Rephrase, Augment, Reason: Visual Grounding of Questions for Vision-Language Models (ICLR 2024 · similarity 0.91)
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.