That could be how it works, but in practice it takes into account all previous tokens when producing the next-token distribution to sample from. So a switch back is more likely than your explanation supposes.