Question for results in DPA2 paper

[duduRUHappy]

Regarding Table 2 in the DPA2 paper: Zero-shot generalization errors, I understand that for all DownStream tasks, DPA2 MT should be a model trained on the entire pre-training dataset. Therefore, for the same downstream task, the results of DPA2 MT should be consistent. However, at the highlighted positions in the figure below, both lines are for AlMgCu-D, but the RMSE for DPA2 MT in terms of Energy and Force are different, with Energy showing a larger deviation between the two.

4.DPA2 Towards a universal large atomic model for molecular and material simulation.pdf

[wanghan-iapcm]

The first zero-shot error was tested with the AlMgCu head while the second was tested with the Alloy head.

[duduRUHappy]

That makes sense now, thank you!