Long-horizon language-conditioned policy learning on CALVIN
98.4Success Rate (Step 5/5)VLA-IAP
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| VLA-IAPToken Retention Ratio=70%, Pruning Ratio=30%2026.03 | 98.4 | 95.1 | 89.7 | 83.7 | 78 | 4.45 | |
| VLA-IAPToken Retention Ratio=50%, Pruning Ratio=50%2026.03 | 98.4 | 94.3 | 89.4 | 83.5 | 77.9 | 4.44 | |
| DivPruneToken Retention Ratio=50%, Pruning Ratio=50%2026.03 | 98.3 | 93.4 | 84.9 | 80.3 | 74.3 | 4.32 | |
| VLA-CacheToken Retention Ratio=70%, Pruning Ratio=30%2026.03 | 98.2 | 94.8 | 89.2 | 82.5 | 77.1 | 4.42 | |
| VLA-CacheToken Retention Ratio=50%, Pruning Ratio=50%2026.03 | 98.1 | 92.5 | 89 | 82.1 | 76.8 | 4.39 | |
| DivPruneToken Retention Ratio=70%, Pruning Ratio=30%2026.03 | 98 | 94.3 | 88.9 | 83.4 | 78.3 | 4.42 | |
| SparseVLMToken Retention Ratio=50%, Pruning Ratio=50%2026.03 | 97.9 | 94.2 | 89.1 | 82.9 | 76.1 | 4.4 | |
| VanillaBase Model=Vanilla, Token Retention Ratio=100%2026.03 | 97.8 | 94.5 | 89.5 | 83.8 | 77.5 | 4.43 | |
| VLA-IAPToken Retention Ratio=30%, Pruning Ratio=70%2026.03 | 97.8 | 94.7 | 89.6 | 83.5 | 76.4 | 4.42 | |
| FastVToken Retention Ratio=70%, Pruning Ratio=30%2026.03 | 97.6 | 94.5 | 89 | 82.8 | 76.8 | 4.4 | |
| SparseVLMToken Retention Ratio=70%, Pruning Ratio=30%2026.03 | 97.4 | 93.8 | 88.6 | 83 | 76.8 | 4.4 | |
| VLA-CacheToken Retention Ratio=30%, Pruning Ratio=70%2026.03 | 97.1 | 90.2 | 82.6 | 78.5 | 68.6 | 4.17 | |
| DivPruneToken Retention Ratio=30%, Pruning Ratio=70%2026.03 | 96.8 | 90.8 | 85 | 76.8 | 69.8 | 4.19 | |
| SparseVLMToken Retention Ratio=30%, Pruning Ratio=70%2026.03 | 96.3 | 90.9 | 83.5 | 74.4 | 66.3 | 4.11 | |
| FastVToken Retention Ratio=50%, Pruning Ratio=50%2026.03 | 94.2 | 92.8 | 84.5 | 79.7 | 70.1 | 4.21 | |
| FastVToken Retention Ratio=30%, Pruning Ratio=70%2026.03 | 89.8 | 85.2 | 78.2 | 71.5 | 63.5 | 3.88 |