Research map / Vision-language and multimodal / Multimodal LLMs
Vision-Language Models Can't See the Obvious
Ngoc Dung Huynh, Phuc H Le-Khac, Wamiq Reyaz Para, Ankit Singh, Sanath Narayan
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 |
|---|---|---|---|
| use synthetic instead of natural images | GPT-4o · SalBench Detection 5-shot | F1: 41.9 -> 86.0 | Improved |
| scale model size from 1.5B to 72B | Qwen2-VL · SalBench Detection synthetic 5-shot | F1: 23.8 -> 54.9 -> 89.9 | Improved |
| scale model size from 1.5B to 72B | Qwen2-VL · SalBench Detection natural 5-shot | F1: 25.3 -> 37.2 -> 43.9 | Improved |
| add few-shot examples (0 to 5 shots) | GPT-4o · SalBench Detection natural | F1: 47.6 -> 38.9 -> 41.9 | Hurt |
| add few-shot examples (0 to 5 shots) | Phi3.5-Vision-3.5B · SalBench Visual Referring synthetic | F1: 63.5 -> 19.0 | Hurt |
| add few-shot examples (0 to 5 shots) | Qwen2-VL-72B · SalBench Visual Referring synthetic | F1: 74.7 -> 87.9 | Improved |
| increase number of distractors (<7 to >25) | Qwen2-VL-72B · SalBench natural Detection | avg F1: 44.5 -> 37.4 | Hurt |
| increase number of distractors (<7 to >25) | GPT-4o · SalBench natural Detection | avg F1: 47.0 -> 41.6 | Hurt |
| increase size difficulty (easy to hard) | Qwen2-VL-72B · SalBench synthetic Size Detection | accuracy: 94.2 -> 46.0 | Hurt |
| increase size difficulty (easy to hard) | GPT-4o · SalBench synthetic Size Detection | accuracy: 93.3 -> 36.8 | Hurt |
| increase color difficulty (medium to hard) | Qwen2-VL-72B · SalBench synthetic Color Detection | accuracy: 100.0 -> 60.1 | Hurt |
| increase orientation difficulty (easy to hard) | Qwen2-VL-72B · SalBench synthetic Orientation Visual Referring | accuracy: 88.3 -> 71.7 | Hurt |
Closest papers by meaning
- Revealing the Gap in Human and VLM Scene Perception through Counterfactual Semantic Saliency (NeurIPS 2026 · similarity 0.89)
- Q-Bench: A Benchmark for General-Purpose Foundation Models on Low-level Vision (ICLR 2024 · similarity 0.89)
- SynDORBench: Evaluating LVLM Perceptual Robustness Under Physically Constrained Visibility Conditions (NeurIPS 2026 · similarity 0.89)
- OddGridBench: Exposing the Lack of Fine-Grained Visual Discrepancy Sensitivity in Multimodal Large Language Models (CVPR 2026 · similarity 0.88)
- MMLongBench: Benchmarking Long-Context Vision-Language Models Effectively and Thoroughly (NeurIPS 2025 · similarity 0.88)
- UniBench: Visual Reasoning Requires Rethinking Vision-Language Beyond Scaling (NeurIPS 2024 · 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.
