Research map / Vision-language and multimodal / Visual question answering
Identifying and Resolving Pitfalls of Knowledge-Based VQA Benchmarks: Auditing, Repairing, and Augmenting
Qian Ma, S M Rayeed, Qiong Wu, Charles Stewart, Yao Ma
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 |
|---|---|---|---|
| repair answers and questions (audit-and-repair protocol) | IBA · InfoSeek All | VQA accuracy: 34.5 -> 42.4 | Improved |
| repair InfoSeek annotations | Wiki-PRF · InfoSeek Time | VQA accuracy: 35.1 -> 43.3 | Improved |
| repair InfoSeek annotations | Wiki-PRF · InfoSeek All | VQA accuracy: 44.9 -> 43.6 | Hurt |
| repair E-VQA questions | IBA · E-VQA | BEM accuracy: 41.9 -> 42.7 | No clear effect |
| fix benchmark: ranking of IBA vs ReflectiVA | IBA vs ReflectiVA · InfoSeek All | VQA accuracy: ReflectiVA 37.3 vs IBA 34.5 before; IBA 42.4 vs ReflectiVA 38.1 after | Mixed |
| question-only fixes on grounded InfoSeek subset (ablation) | EchoSight · InfoSeek shared grounded subset (664) | QA accuracy: 60.1 -> 75.1 | Improved |
| answer-only fixes on grounded InfoSeek subset (ablation) | EchoSight · InfoSeek shared grounded subset (664) | QA accuracy: 44.2 -> 44.3 | No clear effect |
| add single distractor entity (intra-category) | EVA-CLIP-8B retrieval · InfoSeek | Recall@1: 43.5 -> 14.7 | Hurt |
| add single distractor entity (intra-category) | EVA-CLIP-8B retrieval · E-VQA | Recall@1: 13.4 -> 3.5 | Hurt |
| add intra-category distractor | IBA · InfoSeek | VQA accuracy: 40.1 -> 21.4 | Hurt |
| add intra-category distractor | Wiki-PRF · InfoSeek | VQA accuracy: 43.9 -> 23.6 | Hurt |
| post-retrieval evidence selection under augmentation | IBA · InfoSeek post-retrieval recall | recall: 45.1 -> 21.7 | Hurt |
Closest papers by meaning
- StaR-KVQA: Structured Reasoning Traces for Implicit-Knowledge Visual Question Answering (CVPR 2026 · similarity 0.95)
- CRAG-MM-Diagnostics: Enabling Stage-Wise Analysis of Knowledge-Intensive VQA (ECCV 2026 · similarity 0.94)
- OK-VQA: A Visual Question Answering Benchmark Requiring External Knowledge (CVPR 2019 · similarity 0.94)
- Prompting Large Language Models With Answer Heuristics for Knowledge-Based Visual Question Answering (CVPR 2023 · similarity 0.94)
- Separation of Powers: On Segregating Knowledge from Observation in LLM-enabled Knowledge-based Visual Question Answering (CVPR 2025 · similarity 0.93)
- When Open-Vocabulary Visual Question Answering Meets Causal Adapter: Benchmark and Approach (AAAI 2025 · similarity 0.93)
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.
