Visualise models’ hidden thinking tokens