Weight Space Learning without Weights: Leveraging Neural Operators for Model Editing

Weight-space learning enables trained neural networks to be analyzed and modified directly. However, parameter symmetries make the relationship between model weights and represented functions highly non-unique. Data-to-data approaches avoid this parameter ambiguity by operating on sampled signals directly but, in doing so, represent a continuous function through a particular discretization. We explore an alternative approach to model editing that operates directly in function space using neural operators. Focusing on implicit neural representations of MNIST digits, we compare image, function, and weight space approaches for learning geometric transformations. On a fixed $180^\circ$ rotation task, the spatial neural operator achieves the lowest reconstruction error among the tested models, substantially outperforms the weight-space MLP, and remains stable when evaluated at unseen coordinates. Extending the formulation to arbitrary rotations, we find that neural operators generalize to held-out angles. These results provide a proof of concept for function-space approaches to editing neural representations.