Research map / Vision-language and multimodal / Visual question answering
An Empirical Study of GPT-3 for Few-Shot Knowledge-Based 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 |
|---|---|---|---|
| Use PICa-Full with captions and tags instead of the supervised best baseline | OK-VQA · OK-VQA test | accuracy: 39.4 -> 48.0 | Helped |
| Add in-context example selection and multi-query ensemble to PICa-Base (ablation) | PICa · OK-VQA test | accuracy: 42.0 -> 46.9; 43.3 -> 48.0 | Helped |
| Add tags to VinVL captions (ablation) | PICa-Base and PICa-Full · OK-VQA test | accuracy: 42.0 -> 43.3; 46.9 -> 48.0 | Helped |
| Represent the image with text rather than a blank image context (ablation) | PICa · OK-VQA test | accuracy: 24.2 -> 39.3; 30.1 -> 44.6 | Helped |
| Use captions and tags rather than tags alone (ablation) | PICa-Full · OK-VQA test | accuracy: 44.6 -> 48.0 | Helped |
| Use multiple ground-truth captions instead of one ground-truth caption (ablation) | PICa-Full · OK-VQA test | accuracy: 48.7 -> 53.3 | Helped |
| Select examples using question similarity rather than randomly (ablation) | PICa · OK-VQA test | accuracy: 43.3 -> 45.8 | Helped |
| Select dissimilar questions rather than similar questions (ablation) | PICa · OK-VQA test | accuracy: 45.8 -> 40.1 | Hurt |
| Include image similarity with question similarity in example selection (ablation) | PICa-Full · OK-VQA test | accuracy: 45.8 -> 46.5 | Helped |
| Increase the number of ensemble prompts from one to five (ablation) | PICa-Full · OK-VQA test | accuracy: 46.5 -> 48.0 | Helped |
| Use PICa-Full with captions and tags instead of Frozen | VQAv2 · VQAv2 validation | accuracy: 38.2 -> 56.1 | Helped |
| Use text-only image descriptions rather than complete image features | PICa-Full · VQAv2 validation | accuracy: 56.1 -> 73.8 | Hurt |
Closest papers by meaning
- Separation of Powers: On Segregating Knowledge from Observation in LLM-enabled Knowledge-based Visual Question Answering (CVPR 2025 · similarity 0.92)
- BOK-VQA: Bilingual outside Knowledge-Based Visual Question Answering via Graph Representation Pretraining (AAAI 2024 · similarity 0.91)
- When Open-Vocabulary Visual Question Answering Meets Causal Adapter: Benchmark and Approach (AAAI 2025 · similarity 0.90)
- VTQA: Visual Text Question Answering via Entity Alignment and Cross-Media Reasoning (CVPR 2024 · similarity 0.90)
- Bridging the Gap between 2D and 3D Visual Question Answering: A Fusion Approach for 3D VQA (AAAI 2024 · similarity 0.89)
- FOCUS: Internal MLLM Representations for Efficient Fine-Grained Visual Question Answering (NeurIPS 2025 · 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.