Research map / Vision-language and multimodal / Visual question answering
Reliable Visual Question Answering: Abstain Rather Than Answer Incorrectly
Spencer Whitehead, Suzanne Petryk, Vedaad Shakib, Joseph Gonzalez, Trevor Darrell, Anna Rohrbach, Marcus Rohrbach
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 learned Selector instead of MaxProb | CLIP-ViL · VQA v2 test | Coverage at 1% risk: 6.83 -> 15.66 | Improved |
| add learned Selector instead of MaxProb | Pythia · VQA v2 test | Coverage at 1% risk: 5.84 -> 8.30 | Improved |
| add vector-scaling calibration to MaxProb | CLIP-ViL · VQA v2 test | Coverage at 1% risk: 6.83 -> 12.43 | Improved |
| add Selector with high wrong-answer cost (c=100) instead of MaxProb | CLIP-ViL · VQA v2 test | Effective Reliability Phi_100: 2.82 -> 8.00 | Improved |
| allow abstention (Selector) instead of always answering at c=100 | CLIP-ViL · VQA v2 test | Effective Reliability Phi_100: -2162.82 -> 8.00 | Improved |
| add multimodal input to Selector rather than only question (ablation) | Selector on CLIP-ViL · VQA v2 selection validation | Coverage at 1% risk: 0.02 -> 12.92 | Improved |
| regress VQA accuracy instead of classifying correctness (ablation) | Selector on CLIP-ViL · VQA v2 selection validation | Coverage at 1% risk: 6.51 -> 12.92 | Improved |
| use 2-layer MLP instead of 1-layer linear Selector (ablation) | Selector on CLIP-ViL · VQA v2 selection validation | Coverage at 1% risk: 10.38 -> 12.92 | Improved |
| use 4-layer Transformer instead of 2-layer MLP Selector (ablation) | Selector on CLIP-ViL · VQA v2 selection validation | Coverage at 1% risk: 12.92 -> 13.25 | No clear effect |
| use extra Dev data to augment VQA training (MaxProb-Aug) instead of training Selector (ablation) | CLIP-ViL · VQA v2 selection validation | Coverage at 1% risk: 12.92 -> 6.57 | Hurt |
| augment VQA training data with Dev set for MaxProb | CLIP-ViL MaxProb · VQA v2 selection validation | VQA accuracy: 69.70 -> 70.52 | Improved |
Closest papers by meaning
- Improving Selective Visual Question Answering by Learning From Your Peers (CVPR 2023 · similarity 0.94)
- Robust Visual Question Answering: Datasets, Methods, and Future Challenges (TPAMI 2024 · similarity 0.92)
- Beyond Question-Based Biases: Assessing Multimodal Shortcut Learning in Visual Question Answering (ICCV 2021 · similarity 0.92)
- Roses Are Red, Violets Are Blue... but Should VQA Expect Them To? (CVPR 2021 · similarity 0.92)
- When Open-Vocabulary Visual Question Answering Meets Causal Adapter: Benchmark and Approach (AAAI 2025 · similarity 0.91)
- FRAMES-VQA: Benchmarking Fine-Tuning Robustness across Multi-Modal Shifts in Visual Question Answering (CVPR 2025 · similarity 0.90)
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.
