Spatial relationship understanding on MultihopSpatial
37.68AccuracyThinking with Grounding Box
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| Thinking with Grounding BoxModel scale=4B, Grounding interface=Box, Grounding reward=w/2026.06 | 37.68 | 66.4 | |
| Thinking with Grounding PointModel scale=4B, Grounding interface=Point, Grounding reward=w/o2026.06 | 37.03 | 65.4 | |
| Thinking with Grounding PointModel scale=4B, Grounding interface=Point, Grounding reward=w/2026.06 | 37.01 | 65.02 | |
| Thinking with Grounding BoxModel scale=4B, Grounding interface=Box, Grounding reward=w/o2026.06 | 34.89 | 63.82 | |
| Gemma3-27B-ITModel scale=27B2026.06 | 30.94 | 45.82 | |
| Gemma3-12B-ITModel scale=12B2026.06 | 30.08 | 43.58 | |
| Gemma3-4B-ITModel scale=4B2026.06 | 22.7 | 36.87 | |
| Non-grounded ThinkingModel scale=4B, Grounding interface=None2026.06 | 4.79 | 11.67 |