Cross-Layer Discrete Concept Discovery for Interpreting Language Models

Keyphrases

• Concept Vector 20%
• Discrete Latent Space 20%
• Human Annotator 20%
• Jigsaw 20%
• LLM-as-a-Judge 20%
• Map Representation 20%
• Residual Stream 40%
• Single-layer Analysis 20%
• Streaming Features 20%
• VQ-VAE 60%

Engineering

• Discrete Layer 100%
• Layer Analysis 33%

Computer Science

• Adjacent Layer 33%
• Sparse Autoencoder 66%