Research map / Vision-language and multimodal / Multimodal LLMs
Why Do Vision Language Models Struggle To Recognize Human Emotions?
Madhav Agarwal, Sotirios Tsaftaris, Laura Sevilla, Steven McDonagh
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 |
|---|---|---|---|
| shuffle frame order at inference (ablation) | MAE-DFER · MAFW (balanced test, 495 videos) | macro F1: 0.3602 -> 0.3041 | Hurt |
| shuffle frame order at inference (ablation) | HiCMAE · DFEW (balanced test, 700 videos) | macro F1: 0.5725 -> 0.4797 | Hurt |
| shuffle frame order at inference (ablation) | Qwen2.5-VL · MAFW (balanced test, 495 videos) | macro F1: 0.2449 -> 0.2506 | No clear effect |
| shuffle frame order at inference (ablation) | Qwen3-VL · DFEW (balanced test, 700 videos) | macro F1: 0.5511 -> 0.5538 | No clear effect |
| shuffle frame order at inference (ablation) | Gemini2.5-Flash · DFEW (balanced test, 700 videos) | macro F1: 0.6008 -> 0.5778 | Hurt |
| add Multi-Stage Context Enrichment (text summaries of in-between frames) | Qwen2.5-VL · MAFW (balanced test, 495 videos) | macro F1: 0.2449 -> 0.2731 | Improved |
| add MSCE | Qwen2.5-VL · DFEW (balanced test, 700 videos) | macro F1: 0.4552 -> 0.4820 | Improved |
| add MSCE | EmotionQwen · DFEW (balanced test, 700 videos) | macro F1: 0.5010 -> 0.5147 | Improved |
| add MSCE | LLaVA-NeXT-V · MAFW (balanced test, 495 videos) | macro F1: 0.1438 -> 0.1715 | Improved |
| use TCoT frame selection instead of MSCE | Qwen2.5-VL · MAFW (balanced test, 495 videos) | macro F1: 0.245 -> 0.241 | No clear effect |
| add visual prompting (red circle, heatmap, montage, flow arrows) (ablation) | Qwen2.5-VL · MAFW (balanced test, 495 videos) | macro F1: 0.2449 -> 0.1792 (flow arrows, worst) | Hurt |
| increase frame rate beyond 5 FPS (ablation) | Qwen2.5-VL · MAFW | macro F1 | Hurt |
Closest papers by meaning
- Customizing Visual Emotion Evaluation for MLLMs: An Open-vocabulary, Multifaceted, and Scalable Approach (ICLR 2026 · similarity 0.90)
- Catch Your Emotion: Sharpening Emotion Perception in Multimodal Large Language Models (ICML 2025 · similarity 0.89)
- AffectGPT: A New Dataset, Model, and Benchmark for Emotion Understanding with Multimodal Large Language Models (ICML 2025 · similarity 0.89)
- OUS: Bridging Scene Context and Facial Features to Overcome the Rigid Cognitive Problem (AAAI 2025 · similarity 0.87)
- Bridge Then Begin Anew: Generating Target-Relevant Intermediate Model for Source-Free Visual Emotion Adaptation (AAAI 2025 · similarity 0.87)
- OV-MER: Towards Open-Vocabulary Multimodal Emotion Recognition (ICML 2025 · similarity 0.87)
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.
