Research map / Vision-language and multimodal / Visual question answering
CRAG-MM-Diagnostics: Enabling Stage-Wise Analysis of Knowledge-Intensive VQA
Hanseok Oh, Parishad BehnamGhader, Benno Krojer, Hyunji Lee, Paul Pu Liang, Siva Reddy, Verna Dankers
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 disambiguated queries for ambiguous referring expressions (ablation) | visual grounding VLMs · CRAG-MM-Diagnostics grounding | grounding accuracy: mean 44.0% improvement | Improved |
| add disambiguated queries (ablation) | Llama and GPT-5-mini grounding · CRAG-MM-Diagnostics grounding | grounding accuracy | No clear effect |
| crop the target region instead of full image (ablation) | generalized VLMs · Object identification | accuracy | Hurt |
| keep original question instead of simplified prompt (ablation) | generalized VLMs · Object identification | accuracy: 7.9 percentage points | Improved |
| crop the target region before image retrieval (ablation) | specialized image retrievers · Object identification recall@10 | recall@10 | Improved |
| add the question text to multimodal retrievers (ablation) | VLM2Vec2 and Qwen-3-VL-Embedding · Object identification recall@10 | recall@10 | No clear effect |
| provide gold entity name (text-only question) (ablation) | generalized VLMs · KI-VQA accuracy | accuracy: 8.7 points on average | Improved |
| add disambiguated query (ablation) | generalized VLMs · KI-VQA accuracy (ambiguous) | accuracy: 1.8-point gain | Improved |
| add image retrieval (ablation) | Qwen and GPT-5 · KI-VQA accuracy | accuracy: 6.8-point average boost | Improved |
| add text retrieval (ablation) | VLMs · KI-VQA accuracy | accuracy: 6.3 points | Improved |
| use entity-name text-only questions for text retrieval (ablation) | GPT-5 and Qwen · KI-VQA accuracy | accuracy: 16.6 and 26.2 points | Improved |
| add visual grounding crop before image retrieval (ablation) | bimodal RAG · KI-VQA accuracy | accuracy: 13.3 and 8.5 percentage points | Improved |
Closest papers by meaning
- Identifying and Resolving Pitfalls of Knowledge-Based VQA Benchmarks: Auditing, Repairing, and Augmenting (ECCV 2026 · similarity 0.94)
- When Open-Vocabulary Visual Question Answering Meets Causal Adapter: Benchmark and Approach (AAAI 2025 · similarity 0.92)
- VTQA: Visual Text Question Answering via Entity Alignment and Cross-Media Reasoning (CVPR 2024 · similarity 0.92)
- Open-ended VQA benchmarking of Vision-Language models by exploiting Classification datasets and their semantic hierarchy (ICLR 2024 · similarity 0.91)
- Separation of Powers: On Segregating Knowledge from Observation in LLM-enabled Knowledge-based Visual Question Answering (CVPR 2025 · similarity 0.91)
- 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.
