Visual Question Answering on TallyQA (ZOOM-Hard)
16.78GPT AccuracyMags-RL
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Mags-RL2026.05 | 16.78 | 17.27 | |
| Zoom-Refine2026.05 | 10.61 | 7.91 | |
| GRIT2026.05 | 10.61 | 9.35 | |
| Direct Query2026.05 | 9.71 | 7.19 | |
| Chain-of-Thought2026.05 | 7.01 | 5.76 | |
| One-shot ICLICL shots=12026.05 | 5.22 | 3.6 |