Visual Understanding on BLINK sub-tasks
90.67Jigsaw AccuracyInternVL3.5-4B+P^2
Evaluation results
| Method | Links | |||
|---|---|---|---|---|
| InternVL3.5-4B+P^2Protocol=no parameter updates, P2 Plugin Integration=true2026.04 | 90.67 | 86.04 | 62.59 | |
| MirageBackbone=Qwen2.5VL-7B, Protocol=training/fine-tuning2026.04 | 88 | — | — | |
| Qwen3VL-4B+P^2Protocol=no parameter updates, P2 Plugin Integration=true2026.04 | 86.67 | 88.37 | 64.03 | |
| MirageBackbone=Qwen2.5VL-3B, Protocol=training/fine-tuning2026.04 | 85 | — | — | |
| InternVL3.5-2B+P^2Protocol=no parameter updates, P2 Plugin Integration=true2026.04 | 83.33 | 39.53 | 46.04 | |
| InternVL3.5-4BProtocol=no parameter updates, P2 Plugin Integration=false2026.04 | 80 | 65.69 | 45.32 | |
| Zhang et al.Backbone=Qwen2.5VL-7B, Protocol=training/fine-tuning2026.04 | 74.67 | 69.77 | — | |
| Visual SketchpadBackbone=GPT-4 Turbo, Protocol=no parameter updates2026.04 | 68.5 | 52.3 | 42.4 | |
| GPT-4 TurboProtocol=no parameter updates2026.04 | 64.7 | 48.8 | 30.9 | |
| Qwen3VL-4BProtocol=no parameter updates, P2 Plugin Integration=false2026.04 | 64.67 | 86.62 | 61.87 | |
| MMFactoryBackbone=LLaVA-13B, Protocol=no parameter updates2026.04 | 64 | 34.3 | 34.5 | |
| InternVL3.5-2BProtocol=no parameter updates, P2 Plugin Integration=false2026.04 | 61.33 | 34.88 | 33.81 | |
| LLaVA-13BProtocol=no parameter updates2026.04 | 58 | 29.1 | 32.4 | |
| ViGoRLProtocol=training/fine-tuning2026.04 | 56 | — | — |