Metric-semantic goal grounding on MAPG-Bench
42Object Selection AccuracyMAPG
Evaluation results
| Method | Links | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| MAPGBackbone=Gemini 2.5 Pro2026.03 | 42 | 0.08 | 0.45 | 4.9 | 3.8 | 8 | 90 | 94 | 2.5 | |
| SRGPTType=Open-Sourced Specialist2026.03 | 36 | 0.5 | — | — | — | — | — | — | — | |
| MAPGBackbone=Claude Opus 4.62026.03 | 36 | 0.07 | 0.43 | 0.3 | 10.3 | 30 | 98 | 95 | 1.8 | |
| GraphEQAType=Open-Sourced Specialist2026.03 | 34 | — | 5.82 | 13.5 | 27.9 | 14 | 78 | 48 | 1.32 | |
| MAPGBackbone=Gemini 3 Pro2026.03 | 30 | 0.58 | 0.08 | 13.8 | 6.2 | 14 | 46 | 46 | 14.6 | |
| MAPGBackbone=OpenAI GPT 5.22026.03 | 28 | 0.42 | 0.07 | 1.9 | 4.4 | 16 | 98 | 98 | 1.3 |