The Birth of Knowledge: Emergent Features across Time, Space, and Scale in Large Language Models
This addresses the problem of understanding internal representations in LLMs for researchers in mechanistic interpretability, though it is incremental in scope.
The paper investigates the emergence of interpretable categorical features in large language models across training time, layers, and model sizes, finding clear thresholds for feature emergence and unexpected semantic reactivation patterns that challenge standard assumptions.
This paper studies the emergence of interpretable categorical features within large language models (LLMs), analyzing their behavior across training checkpoints (time), transformer layers (space), and varying model sizes (scale). Using sparse autoencoders for mechanistic interpretability, we identify when and where specific semantic concepts emerge within neural activations. Results indicate clear temporal and scale-specific thresholds for feature emergence across multiple domains. Notably, spatial analysis reveals unexpected semantic reactivation, with early-layer features re-emerging at later layers, challenging standard assumptions about representational dynamics in transformer models.