Point Grounding on PointArena Point-Bench (cross-dataset evaluation)
83.8Affordance P-AccQwen2.5-VL-7B Ours
Evaluation Results
| Method | Links | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Qwen2.5-VL-7B OursBase MLLM=Qwen2.5-VL-7B, Method=Ours (Q-Synth + A2P Decoder)2026.06 | 83.8 | 82.8 | 58.5 | 54.4 | 65.8 | 62.2 | 43 | 39 | |
| Molmo-7B OursBase MLLM=Molmo-7B, Method=Ours (Q-Synth + A2P Decoder)2026.06 | 79.3 | 74.8 | 55.4 | 50.8 | 65.3 | 59.1 | 45 | 41 | |
| First-Gen-MLLM OursBase MLLM=First-Gen-MLLM, Method=Ours (Q-Synth + A2P Decoder)2026.06 | 67.2 | 64.1 | 42.4 | 38.4 | 43.9 | 37.8 | 40.5 | 36.9 | |
| First-Gen-MLLM attention pointingBase MLLM=First-Gen-MLLM, Method=native attention pointing2026.06 | 51 | 41.7 | 25.6 | 23.2 | 38.9 | 32.2 | 17.3 | 16.7 | |
| Molmo-7B attention pointingBase MLLM=Molmo-7B, Method=native attention pointing2026.06 | 49.5 | 41.9 | 43.6 | 36.9 | 51.3 | 43 | 28.5 | 27.5 | |
| Qwen2.5-VL-7B attention pointingBase MLLM=Qwen2.5-VL-7B, Method=native attention pointing2026.06 | 35.9 | 30.3 | 43.1 | 38 | 37.3 | 32.1 | 22.7 | 17.2 |