Long-horizon Embodied Task on ∞-THOR 1.0 (online evaluation)
6.2Go toMemory-Augmented
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Memory-AugmentedInput modality=Image, Top-k=202025.05 | 6.2 | 0 | 0 | 3.1 | |
| Memory-AugmentedInput modality=Text2025.05 | 10.8 | 1.5 | 0 | 6.1 | |
| Interleaved Goal-State-ActionFine-tuning context size=32K2025.05 | 11.3 | 1.5 | 0 | 6.4 | |
| Interleaved Goal-State-ActionFine-tuning context size=128K2025.05 | 18.5 | 2.3 | 0 | 9.2 | |
| Interleaved Goal-State-ActionFine-tuning context size=128K, Inference Extension=Dynamic Scaling2025.05 | 20 | 2.3 | 0 | 9.9 | |
| Interleaved Goal-State-ActionFine-tuning context size=128K, Inference Extension=YaRN (x4)2025.05 | 23.1 | 2.6 | 0 | 11.5 |