Logit Lens Analysis

A neural network interpretability method that directly projects the hidden states of each Transformer layer into the vocabulary logit space, tracking how model predictions evolve across layers to reveal how output decisions are formed during the forward pass.