One way of looking at it (not thinking too hard about the differences between TE, TM, and hybrid modes) is to consider the momentum vector, or k-vector, of the mode.
Like any vector, a k-vector can be decomposed as \$\vec{k}=k_x \hat{x} + k_y \hat{y} + k_z \hat{z}\$. If the waveguide is aligned in the z-direction, then \$k_x\$ and \$k_y\$ tell us how much of the wave's momentum is (speaking crudely) bouncing back and forth between the walls of the waveguide, and \$k_z\$ tells us how much of the momentum is devoted to propagating the wave along the guide.
The boundary conditions for the (n,m)'th mode to propagate can be expressed as
$$k_x=\frac{n\pi}{\ell_x}$$
and
$$k_y=\frac{m\pi}{\ell_y}$$
As the waveguide dimensions \$\ell_x\$ and \$\ell_y\$ shrink, more of the k-vector must be used to fulfill these boundary conditions and less is available for the \$k_z\$ component.
When \$|\vec{k}|^2 = k_x^2 + k_y^2\$, and therefore \$k_z=0\$, we have reached the cut-off condition.
Basically we have a wave that is resonating in the x and y directions (transverse to the guide) and simply a plane wave (no variation) in the z direction.