What problem does it solve? Understanding what information is encoded inside a large language model's internal states is difficult because hidden vectors are not human-readable. Vocabulary Projection decodes residual stream states, attention head outputs, and FFN neuron value weights into distributions over the model's vocabulary, letting researchers read semantic content directly from intermediate computations without training probes or collecting labeled datasets. ## Core Features & Use Cases - Residual Stream Projection: Trace layer-by-layer prediction evolution with tuned lens analysis to identify where specific concepts emerge in the network. - Attention Head Output Projection: Decode individual head outputs to reveal what information (e.g., copied names, next-token candidates) each head transmits, supporting functional head identification. - Neuron Value Weight Projection: Treat FFN layers as key-value memories and project value vectors to see which token clusters individual neurons promote. - Use Case: A mechanistic interpretability researcher analyzing Llama-3.1-8B can apply logit lens analysis across all layers, visualize confidence heatmaps per generation step, and pinpoint the layers where a factual prediction crystallizes. ## Quick Start Ask the agent to run a logit lens analysis on a prompt with your chosen model and visualize how the predicted tokens evolve across transformer layers.
