Visual Question Answering (Exact Match Accuracy) on ScienceQA
86.37Exact Match AccuracyGemma-3-4b
Evaluation Results
| Method | Links | |
|---|---|---|
| Gemma-3-4bTraining Stage=SCALe-GRPO2026.03 | 86.37 | |
| Gemma-3-4bTraining Stage=SCALe-SFT2026.03 | 85.92 | |
| Qwen2.5-VL-3BTraining Stage=SCALe-GRPO2026.03 | 85.82 | |
| Gemma-3-4bTraining Stage=SFT2026.03 | 85.77 | |
| Qwen2.5-VL-3BTraining Stage=SCALe-SFT2026.03 | 85.52 | |
| Gemma-3-4bTraining Stage=GRPO2026.03 | 85.18 | |
| Qwen2.5-VL-3BTraining Stage=GRPO2026.03 | 85.08 | |
| Qwen2.5-VL-3BTraining Stage=SFT2026.03 | 85 | |
| LLaVA-v1.6-Mistral-7BTraining Stage=SCALe-GRPO2026.03 | 83.39 | |
| LLaVA-v1.6-Mistral-7BTraining Stage=SCALe-SFT2026.03 | 82.6 | |
| LLaVA-v1.6-Mistral-7BTraining Stage=GRPO2026.03 | 80.86 | |
| Qwen2.5-VL-3BTraining Stage=Base2026.03 | 80.6 | |
| LLaVA-v1.6-Mistral-7BTraining Stage=SFT2026.03 | 80.17 | |
| Gemma-3-4bTraining Stage=Base2026.03 | 77.68 | |
| LLaVA-v1.6-Mistral-7BTraining Stage=Base2026.03 | 72.58 |