Looks good to me. Have tested against my failing minimal reproducible example and it's fixed the issue.

I don't know how feasible it is, but adding an extra test case with a predictor that has uncertainty built in might have caught this issue?

We already have test cases for MVE and evidential methods (see here as an example). To capture the issue you found, we would need to test regression tasks without target normalization. AFAIK, we don't have a flag to disable target scaling, since the target should always be scaled to ensure stable model training (we also scale targets in the test you refer to).

@shihchengli We scale the targets in the test he refers to, but we don't use the scaler from that to make an unscale transform for the FFN.
In conftest we have both a normal mpnn and a dirichlet mpnn defined, both without an unscale transform in the FFN.

We use the classification_mpnn_dirichlet in integration/test_classification_mol.py

We could do something similar and have an MVE mpnn defined in conftest.py and use it in integration/test_regression_mol.py. Then we would test if MVE works without the unscale transform and the CLI tests will still test if MVE works with an unscale transform.