Backward Lens: Projecting Language Model Gradients into the Vocabulary Space

Understanding how Transformer-based Language Models (LMs) learn and recall information is a key goal of the deep learning community. Recent interpretability methods project weights and hidden states obtained from the forward pass to the models' vocabularies, helping to uncover how information flows within LMs. In this work, we extend this methodology to LMs' backward pass and gradients. We first prove that a gradient matrix can be cast as a low-rank linear combination of its forward and backward passes' inputs. We then develop methods to project these gradients into vocabulary Figure 1: An illustration depicting the tokens promoted items and explore the mechanics of how by a single LM's MLP layer and its gradient during the new information is stored in the LMs' neurons.