What hidden states should I use?

[Bird.Z]

the output of encoder is N*d states for each text input;

1. simcse use the 1st one as the output.
2. the sentence-transformers use the mean pooling(only for attention mask=1 states).

Your default is the same as mean pooling?

[yuxin.wang]

yes, use the mean pooling(only for attention mask=1 states)