Why are all the predicted outputs the same?

I run the code on Multi-Xscience with the recommended setup. But I find all the predicted outputs on the test set are the same, as follows:

Do you have any idea about the reason? Thx.

Start by reproducing the recommended setup on the Multi-Xscience test set and inspect the prediction-generation path and configuration used for inference. Compare the inputs and decoding behavior across examples; done means the cause of identical outputs is identified and predictions are no longer all the same.

Written by the indexing model from the issue text.