Research map / Vision-language and multimodal / Visual question answering
Re-evaluating Continual VQA: Toward Fair and Robust Evaluation for Multimodal Continual Learning
Zijian Gao, Zicheng Sun, Xingxing Zhang, Kele Xu, Huaimin Wang
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 answer prediction distillation (LAPD) with question-only replay (ablation) | MaDQ · GQA v2 (PS) | FAA: 4.98 -> 36.88 | Improved |
| add image-question matching loss (LIQM) on top of LAPD (ablation) | MaDQ · GQA v2 (PS) | CAA: 41.75 -> 43.36 | Improved |
| add matching consistency distillation (LMCD) on top of LAPD and LIQM (ablation) | MaDQ · GQA v2 (PS) | FFM: 11.07 -> 9.37 | Improved |
| add LMCD on top of LAPD and LIQM (ablation) | MaDQ · VQA v3 (PS) | FFM: 6.83 -> 4.54 | Improved |
| use LIQM without any distillation loss (ablation) | MaDQ · GQA v2 (PS) | FAA: 4.98 -> 5.06 | No clear effect |
| plug LIQM into existing continual learning methods (ablation) | LwF, ER, CLS-ER · GQA v2 (PS) and VQA v3 (PS) | FFM: CLS-ER 11.87 -> 10.71 (GQA v2); 7.56 -> 5.46 (VQA v3) | Improved |
| replace full-triplet replay (ER, CLS-ER) with question-only replay plus dual-level distillation (MaDQ) | ER / CLS-ER · VQA v3 (PS) | CAA: 33.64 -> 37.82 | Improved |
| MaDQ vs ER/CLS-ER under standard splits on GQA v2 | ER / CLS-ER · GQA v2 (SS) | FAA: 41.61 -> 43.51 | Improved |
| use token-level disjoint answer spaces (GQA v2 vs original GQA) | continual VQA benchmarks · GQA vs GQA v2 (SS) | FAA / CAA / FFM | Hurt |
| apply simplified MaDQ* to BLIP2 instead of BLIP | MaDQ · VQA v3 (SS) | FAA: 21.48 (MaDQ*) vs 23.46 (ER) and 24.45 (CLS-ER) | Mixed |
| use BLIP2 LLM-based backbone instead of BLIP | continual VQA methods · VQA v3 (SS) | FAA / CAA | Hurt |
| introduce intra-task train-test answer shift (PS) versus standard splits | continual VQA methods · VQA v3 | FAA: MaDQ 53.21 -> 36.15; Joint 69.41 -> 42.90 | Hurt |
Closest papers by meaning
- VQACL: A Novel Visual Question Answering Continual Learning Setting (CVPR 2023 · similarity 0.95)
- Ask and Remember: A Questions-Only Replay Strategy for Continual Visual Question Answering (ICCV 2025 · similarity 0.94)
- MacVQA: Adaptive Memory Allocation and Global Noise Filtering for Continual Visual Question Answering (AAAI 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 0.91)
- Dynamic Cross-Modal Prompt Generation for Multimodal Continual Instruction Tuning (NeurIPS 2026 · similarity 0.88)
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.
