Research map / Vision-language and multimodal / Visual question answering
Ask and Remember: A Questions-Only Replay Strategy for Continual Visual Question Answering
Imad Eddine Marouf, Enzo Tartaglione, Stéphane Lathuilière, Joost Van De Weijer
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 attention consistency distillation to question-only replay (ablation) | QUAD · VQAv2 (standard test) | AP: 30.72 -> 39.25 | Improved |
| add attention consistency distillation to question-only replay (ablation) | QUAD · NExT-QA (standard test) | AP: 29.04 -> 31.70 | Improved |
| add question-only replay to attention distillation (ablation) | QUAD · VQAv2 (standard test) | AP: 13.34 -> 39.25 | Improved |
| add question-only replay to attention distillation (ablation) | QUAD · NExT-QA (standard test) | AP: 13.24 -> 31.70 | Improved |
| add attention distillation to question-only replay (ablation) | QUAD · VQAv2 (standard test) | Forgetting: 13.74 -> 4.91 | Improved |
| use cross-entropy on softmax attention instead of L1 on raw attention (ablation) | QUAD · VQAv2 (standard test) | AP: 34.56 -> 39.25 | Improved |
| use cross-entropy on softmax attention instead of L1 on raw attention (ablation) | QUAD · NExT-QA (standard test) | AP: 30.14 -> 31.70 | Improved |
| use cross-entropy attention distillation instead of asymmetric ReLU+L1 (ablation) | QUAD · VQAv2 (standard test) | AP: 38.15 -> 39.25 | Improved |
| use cross-entropy attention distillation instead of asymmetric ReLU+L1 (ablation) | QUAD · NExT-QA (standard test) | Forgetting: 4.13 -> 2.91 | Improved |
| replace stored image-question pairs with question-only replay (no images) | QUAD vs VQACL · VQAv2 (standard test) | AP: 37.46 -> 39.25 | Improved |
| replace stored image-question pairs with question-only replay (no images) | QUAD vs VQACL · NExT-QA (standard test) | AP: 30.86 -> 31.70 | Mixed |
| question-only replay vs image storage on pretrained LLaVA-7B | QUAD vs CL-MoE · VQAv2 (LLaVA-7B pretrained) | AP: 52.96 -> 44.16 | Hurt |
Closest papers by meaning
- VQACL: A Novel Visual Question Answering Continual Learning Setting (CVPR 2023 · similarity 0.94)
- Re-evaluating Continual VQA: Toward Fair and Robust Evaluation for Multimodal Continual Learning (CVPR 2026 · similarity 0.94)
- Attention-Logit Steering to Compositional Generalization for Continual VQA (ECCV 2026 · similarity 0.94)
- AVQACL: A Novel Benchmark for Audio-Visual Question Answering Continual Learning (CVPR 2025 · similarity 0.91)
- CL-MoE: Enhancing Multimodal Large Language Model with Dual Momentum Mixture-of-Experts for Continual Visual Question Answering (CVPR 2025 · similarity 0.90)
- Language-aware Visual Semantic Distillation for Video Question Answering (CVPR 2024 · 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.
