Meme Classification on ArMeme (test)
76.6AccuracyQwen3-VL-8B-Instruct
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Qwen3-VL-8B-InstructClassification Head Configuration=Seq. cls. baseline2026.06 | 76.6 | 75.3 | 59.4 | |
| Gemma-3-12B-ITClassification Head Configuration=Seq. cls. baseline2026.06 | 74.8 | 72.4 | 59.7 | |
| Proposed (+ Self-Sup.)Training Protocol=Self-supervised GRPO, Inference Strategy=Reasoning-based2026.06 | 72.8 | 71.5 | 61.2 | |
| Proposed (Sup.)Training Protocol=Supervised GRPO, Inference Strategy=Reasoning-based2026.06 | 72.6 | 71.1 | 59.7 | |
| Llama MSSetup=FT, Optimization=Multi-stage2025.02 | 72.1 | 69.9 | 53.6 | |
| [43]Context=Prior work2026.06 | 72.1 | 69.9 | 53.6 | |
| Llama (Ar-Exp) MSSetup=FT, Optimization=Multi-stage, Training=Arabic explanation2025.02 | 72 | 69.6 | 49.9 | |
| mBERTSetup=mBERT2025.02 | 70.7 | 67.5 | 48.7 | |
| QaribSetup=Qarib2025.02 | 69.7 | 69 | 55.1 | |
| [5]Context=Prior work2026.06 | 69.7 | 69 | 55.1 | |
| ResNet50Setup=ResNet502025.02 | 66 | 63.7 | 43.4 |