Current models simply generate additional text that gets added to the context for the trace. However iterative models that "think" by repeatedly looping through several layers instead of outputting text have recently been demonstrated.