Visual Grounding on InsPLAD
100Parse ScoreGemma 4 12B
Evaluation Results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Gemma 4 12BSetting=+ prompt budget 8, Prompt Budget=82026.06 | 100 | 5.025 | 49.4 | 38.1 | 2.1 | 8 | |
| Gemma 4 12BSetting=+ budget 8 + repeat-stop, Prompt Budget=8, Mechanism=object-level repeat-stop2026.06 | 100 | 4.787 | 49 | 38.5 | 0 | 1 | |
| Qwen3-VL-8BSetting=+ rep penalty 1.05, Repetition Penalty=1.052026.06 | 100 | 3.837 | 24.9 | 21.4 | 0.2 | 2 | |
| Qwen3-VL-8BSetting=+ budget 6 + repeat-stop, Prompt Budget=6, Mechanism=object-level repeat-stop2026.06 | 100 | 3.475 | 25.9 | — | 0 | 1 | |
| Qwen3-VL-8BSetting=q/v r32 + budget 6 + stop, Rank=32, Prompt Budget=6, Mechanism=object-level repeat-stop2026.06 | 100 | 4.838 | 31.8 | 27.5 | 0 | 1 | |
| Qwen3-VL-8BSetting=Base2026.06 | 98.8 | 2.925 | 15.5 | 14.3 | 1.6 | 5 | |
| Qwen3-VL-8BSetting=q/v r8, Fine-tuning=Dense Coordinate-List2026.06 | 98.8 | 5.013 | 23.3 | 18.6 | 8.1 | 12 | |
| Qwen3-VL-8BSetting=+ prompt budget 6, Prompt Budget=62026.06 | 98.8 | 3.487 | 25.8 | — | 4.4 | 6 | |
| Gemma 4 12BSetting=Base2026.06 | 96.3 | 4.763 | 0.7 | 0.2 | 0.2 | 2 | |
| Gemma 4 12BSetting=+ rep penalty 1.05, Repetition Penalty=1.052026.06 | 90 | 6.425 | 47.4 | 36.6 | 3.9 | 19 | |
| Gemma 4 12BSetting=q/k/v/o r32, Fine-tuning=Dense Coordinate-List2026.06 | 81.2 | 6.713 | 44.8 | 36.3 | 8 | 23 |