Comment by wlib

The filler tokens actually do make them think more. Even just allowing the models to output "." until they are confident enough to output something increases their performance. Of course, training the model to do this (use pause tokens) on purpose works too: https://arxiv.org/pdf/2310.02226