Research map / Vision-language and multimodal / Visual question answering
How Transferable Are Reasoning Patterns in VQA?
Check 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 |
|---|---|---|---|
| transfer oracle parameters and fine-tune on noisy visual inputs | tiny-LXMERT · GQA-OOD-testdev | acc-tail: 42.9 -> 48.5 | Helped |
| transfer oracle parameters and fine-tune on noisy visual inputs | tiny-LXMERT · GQA-testdev | overall accuracy: 52.4 -> 56.8 | Helped |
| add Oracle Transfer to LXMERT/BERT-pretrained model | tiny-LXMERT with LXMERT/BERT pretraining · VQAv2-test-std | overall accuracy: 69.7 -> 70.2 | Helped |
| add LXMERT/BERT pretraining to Oracle Transfer | tiny-LXMERT with Oracle Transfer · GQA-testdev | overall accuracy: 56.8 -> 57.8 | Helped |
| transfer oracle weights without retraining, using predicted 1-in-K classes instead of dense inputs (ablation) | oracle model · GQA-val | accuracy: 61.7 -> 58.8 | Hurt |
| retrain only the visual block T_V after oracle transfer (ablation) | oracle model · GQA-val | accuracy: 61.7 -> 61.7 | No clear effect |
| retrain the full model on dense visual inputs after oracle transfer (ablation) | oracle model · GQA-val | accuracy: 61.7 -> 66.3 | Helped |
| remove language self-attention by replacing its attention maps with uniform attention (ablation) | oracle VL-Transformer · GQA-val | accuracy: 91.5 -> 37.9 | Hurt |
| remove vision self-attention by replacing its attention maps with uniform attention (ablation) | oracle VL-Transformer · GQA-val | accuracy: 91.5 -> 91.4 | No clear effect |
| remove language-to-vision cross-attention by replacing its attention maps with uniform attention (ablation) | oracle VL-Transformer · GQA-val | accuracy: 91.5 -> 52.8 | Hurt |
| remove vision-to-language cross-attention by replacing its attention maps with uniform attention (ablation) | oracle VL-Transformer · GQA-val | accuracy: 91.5 -> 68.1 | Hurt |
Closest papers by meaning
- Debiased Visual Question Answering from Feature and Sample Perspectives (NeurIPS 2021 · similarity 0.93)
- Roses Are Red, Violets Are Blue... but Should VQA Expect Them To? (CVPR 2021 · similarity 0.93)
- Making the v in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering (CVPR 2017 · similarity 0.92)
- SQuINTing at VQA Models: Introspecting VQA Models With Sub-Questions (CVPR 2020 · similarity 0.92)
- GQA: A New Dataset for Real-World Visual Reasoning and Compositional Question Answering (CVPR 2019 · similarity 0.92)
- When Open-Vocabulary Visual Question Answering Meets Causal Adapter: Benchmark and Approach (AAAI 2025 · 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.