Research map / Vision-language and multimodal / Visual question answering
Beyond Question-Based Biases: Assessing Multimodal Shortcut Learning in Visual Question Answering
Corentin Dancette, Rémi Cadène, Damien Teney, Matthieu Cord
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 LfF debiasing | UpDown · VQA v2 Counterexamples | accuracy: 33.91 -> 34.27 | Improved |
| add RandImg debiasing | UpDown · VQA v2 Counterexamples | accuracy: 33.91 -> 34.41 | Improved |
| add LMH debiasing | UpDown · VQA v2 Counterexamples vs Easy | accuracy: 33.91 -> 34.26 (Counterexamples), 76.69 -> 73.12 (Easy) | Mixed |
| add LMH+CSS debiasing | UpDown · VQA v2 Counterexamples vs Easy | accuracy: 33.91 -> 34.36 (Counterexamples), 76.69 -> 62.08 (Easy) | Mixed |
| add RUBi debiasing | UpDown · VQA v2 Counterexamples | accuracy: 33.91 -> 32.25 | Hurt |
| add LMH+RMFE debiasing | UpDown · VQA v2 overall | accuracy: 63.52 -> 60.96 | Hurt |
| add ESR debiasing | UpDown · VQA v2 overall | accuracy: 63.52 -> 62.96 | Hurt |
| use BLOCK bilinear fusion instead of UpDown | UpDown · VQA v2 Counterexamples | accuracy: 33.91 -> 32.91 | Hurt |
| use BLOCK bilinear fusion instead of UpDown | UpDown · VQA v2 Easy | accuracy: 76.69 -> 77.65 | Improved |
| add VilBERT pretraining on Conceptual Captions | UpDown-style models · VQA v2 Counterexamples | accuracy: 33.91 -> 39.24 | Improved |
| move from overall to Counterexamples subset | VQA models SAN, UpDown, BLOCK, VilBERT · VQA v2 | accuracy: about 29 points lost | Hurt |
| apply bias-reduction methods trained for VQA-CP to natural VQA v2 shortcuts | UpDown · VQA-CP v2 vs VQA-CE | accuracy: +5 to +20 on VQA-CP, gains of +0.36 and +0.50 at best on Counterexamples | Mixed |
Closest papers by meaning
- Roses Are Red, Violets Are Blue... but Should VQA Expect Them To? (CVPR 2021 · similarity 0.94)
- Debiased Visual Question Answering from Feature and Sample Perspectives (NeurIPS 2021 · similarity 0.94)
- Making the v in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering (CVPR 2017 · similarity 0.93)
- An Analysis of Visual Question Answering Algorithms (ICCV 2017 · similarity 0.93)
- When Open-Vocabulary Visual Question Answering Meets Causal Adapter: Benchmark and Approach (AAAI 2025 · similarity 0.92)
- FOCUS: Internal MLLM Representations for Efficient Fine-Grained Visual Question Answering (NeurIPS 2025 · 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.
