Microsoft and Cornell researchers show pause tokens can buy models more inference compute

The research explores a token that lets a model spend additional sequential computation before producing the next substantive output.