Comment by KTibow

This can still happen even with thinking models as long as the model outputs tokens in a sequence. Only way to fix would be to allow it to restart its response or switch to diffusion.