[R] Ternary neural networks as a path to more efficient AI - is (+1, 0, -1) weight quantization getting serious research attention?

I've been reading about ternary weight quantization in neural networks and wanted to get a sence of how seriously the ML research community is taking this direction.The theoretical appeal seems clear: ternary weights (+1, 0, -1) cut model size and inference cost a lot compared to full-precision or even binary networks, while keeping more power than strict binary. Papers like TWN (Ternary Weight Networks) from 2016 and some newer work suggest this is a real path for efficient inference.What I've been less clear on is the training story. Most ternary network research I've seen focuses on post-training quantization - you train in full precision and then quantize. But I came across a reference to an architecture that claims to train natively in ternary, using an evolutionary selection mechanism rather than gradient descent.The claim is that native ternary training produces models that represent uncertainty more naturally and stay adaptive rather than freezing after training. The project is called Aigarth, developed by Qubic.I'm not in a position to evaluate the claim rigourously. But the combination of native ternary training + evolutionary optimization rather than backpropagation is unusual enough that I wanted to ask: is this a known research direction? Are there peer-reviewed papers exploring native ternary training with evolutionary methods? Is this genuinely novel or am I missing obvious prior work?