I suspect it works as regularisation of the network. It usually happens when you train with quantisation instead of post-training quantisation, an I haven't see

I suspect it works as regularisation of the network. It usually happens when you train with quantisation instead of post-training quantisation, an I haven't seen that done with LLMs yet.