The NEF allows you to use almost any neuron model as long as it has an equation for it's activity and it's spikes in some way. Usually, a simple leaky-integrate-fire (LIF) neuron model is used, but even in the software Nengo there are various neurons available for modelling, such as adaptive LIF neurons and Poisson neurons. What are the benefits of using these more complex neurons and how have these benefits been leveraged in the past?
As seen in "Neural Engineering" by Eliasmith et al. Chapter 4, complicated neuron models have greater computational abilities and match neural data more realistically.
As seen in the following table (taken from Neural Engineering) Adaptive LIF neurons, by virtue of their temporally varying firing patterns (i.e. a constant input will cause different spike patterns over time) encode more information per spike:
As seen in this table, there's a trade-off between the complexity of the information represented by the neuron and it's computational cost. However, this cost is calculated on traditional Von-Neumann CPUs. It is possible to create neuromorphic hardware to minimize this cost while maximizing computational ability. For more information, see "Nonlinear synaptic interaction as a computational resource in the Neural Engineering Framework" which shows how non-linear dendrites can allow for powerful forms of computation with minimal additional neuromorphic hardware cost.
Matching Biological Data
The aforementioned temporal variance of the Adaptive LIF is also used to matching neural data as seen in this model of working memory, which would have not been possible with regular LIF neurons. Using even more complicated models can allow for matching the effects of drugs, as shown in "Effects of Guanfacine and Phenylephrine on a Spiking Neuron Model of Working Memory".