💡 This paper proposes differentiable logic gate networks that combine real-valued logic and continuous parameter relaxation to enable gradient-based training of non-differentiable logic gate networks, a