Hi, thank you for this work.
Was there an ablation done to measure the performance of the trained model (pretrained on spatial datasets) but without any test-time updates?
I tried this with the released nano checkpoint and performance was around the same as with TTT, which seems surprising as it suggests that the gradient steps at test-time do not contribute much to the performance.
Hi, thank you for this work.
Was there an ablation done to measure the performance of the trained model (pretrained on spatial datasets) but without any test-time updates?
I tried this with the released nano checkpoint and performance was around the same as with TTT, which seems surprising as it suggests that the gradient steps at test-time do not contribute much to the performance.