Feature-Level Interpretability at Frontier Scale

Abstract

We scale feature-level interpretability methods to frontier-size models and report what fraction of computation the resulting features account for. Coverage is partial. We discuss what level of coverage would be required before interpretability could support a decision to remove human review from a deployment, and note that current coverage is far from it.

Key findings

• 01Methods operate at frontier scale without prohibitive cost
• 02Coverage is partial; most computation remains uninspected
• 03Features recovered are stable across training runs
• 04Coverage sufficient for oversight decisions is not close

Limitations

Published at equal prominence to the findings. A paper’s limitations are usually the part that determines how much its result should move your beliefs.

• Single model family
• Coverage measurement depends on contested definitions
• No causal validation of the recovered features