Forward Self-Models Learn an Empirical Approximation of Neural Network Computation

We introduce forward self-models: small networks trained to predict a neural network's later-layer activations from its earlier-layer activations, learning an empirical approximation of the computational function that the intervening layers implement. We demonstrate this by training forward self-models with sizes of 1-3% of main model parameters on main models ranging from 30M to 1B parameters, achieving up to 97% cosine similarity with the target activations and up to 94% recovery of the KL divergence of a layer's contribution to the output distribution. Forward self-model prediction errors are interpretably structured and track the computational complexity gap between the main model and forward model. We argue that forward self-models provide a primitive for both mechanistic interpretability and future model architectures that require an explicit model of their own computational dynamics.