Research map / Vision-language and multimodal / Visual question answering
VisualOverload: Probing Visual Understanding of VLMs in Really Dense Scenes
Paul Gavrikov, Wei Lin, M. Jehanzeb Mirza, Soumya Jahagirdar, Muhammad Huzaifa, Sivan Doveh, James Glass, Serena Yeung-Levy, Hilde Kuehne
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 zero-shot chain-of-thought prompting (ablation) | InternVL3 38B · VisualOverload total | accuracy: 67.6 -> 65.5 | Hurt |
| add zero-shot chain-of-thought prompting (ablation) | InternVL3 38B · VisualOverload hard split | accuracy: 7.2 -> 14.4 | Improved |
| add zero-shot chain-of-thought prompting (ablation) | InternVL3 38B · VisualOverload OCR | accuracy: 45.8 -> 50.0 | Improved |
| add zero-shot chain-of-thought prompting (ablation) | InternVL3 38B · VisualOverload reasoning | accuracy: 69.8 -> 62.4 | Hurt |
| scale InternVL3 from 38B to 78B | InternVL3 · VisualOverload total | accuracy: 67.6 -> 66.8 | Hurt |
| scale PaliGemma 2 from 10B to 28B | PaliGemma 2 · VisualOverload total | accuracy: 50.3 -> 41.5 | Hurt |
| scale Qwen2.5-VL from 3B to 7B | Qwen2.5-VL · VisualOverload total | accuracy: 54.1 -> 61.5 | Improved |
| use specialized high-resolution model instead of regular model | VILA HD 4K vs Qwen2.5-VL 7B · VisualOverload total | accuracy: 61.5 -> 48.5 | Hurt |
| allow 10% tolerance in counting answers | all evaluated VLMs · VisualOverload counting | accuracy: +1.6% | No clear effect |
| move from scene classification to reasoning questions (logical consistency) | frontier VLMs · VisualOverload | logical consistency (%): 83.3 -> 60.6 | Hurt |
| increase image resolution up to Full HD (ablation) | InternVL3 · VisualOverload | accuracy | Improved |
Closest papers by meaning
- OK-VQA: A Visual Question Answering Benchmark Requiring External Knowledge (CVPR 2019 · similarity 0.92)
- Making the v in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering (CVPR 2017 · similarity 0.92)
- A-OKVQA: A Benchmark for Visual Question Answering Using World Knowledge (ECCV 2022 · similarity 0.92)
- FOCUS: Internal MLLM Representations for Efficient Fine-Grained Visual Question Answering (NeurIPS 2025 · similarity 0.91)
- When Open-Vocabulary Visual Question Answering Meets Causal Adapter: Benchmark and Approach (AAAI 2025 · similarity 0.90)
- Bridging the Gap between 2D and 3D Visual Question Answering: A Fusion Approach for 3D VQA (AAAI 2024 · similarity 0.90)
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.
