SAE Feature Interpretability Improves Across Layers

Sauers_ · x · 2026-07-14

This repost discusses how jlens is used to describe SAE features, essentially verbalizing highly correlated features like "motor neurons."

By comparing the consistency between jlens-generated feature descriptions and Neuronpedia labels, the author observed that this "verbalization" capability improves with depth. In Qwen3-4b, a significant shift occurs around layer 22. The original post also notes a similar trend in the raw Jacobian results, where later layers exhibit higher Jacobian gains.

Original post →

More from Research

Research channel →