Visual-Spatial Reasoning on SceneGraphs (test)
83.82AccuracyNAG-LoRA
Evaluation results
| Method | Links | |
|---|---|---|
| NAG-LoRABackbone=Qwen3-600M, Evaluation Protocol=LoRA Fine-tuning2026.01 | 83.82 | |
| Qwen3-LoRABackbone=Qwen3-600M, Evaluation Protocol=LoRA Fine-tuning2026.01 | 83.71 | |
| GraphTokenNBackbone=Qwen3-600M, Granularity=Node-level2026.01 | 73.65 | |
| GraphTokenEBackbone=Qwen3-600M, Granularity=Edge-level2026.01 | 72.66 | |
| NAG-ZeroBackbone=Qwen3-600M, Evaluation Protocol=Lightweight Adapter Training2026.01 | 71.51 | |
| GraphTokenGBackbone=Qwen3-600M, Granularity=Graph-level2026.01 | 68.37 | |
| Qwen3-DirectBackbone=Qwen3-600M, Evaluation Protocol=Direct Text-only Inference2026.01 | 50 |