Autointerp Simulation Scores Track Token Naming, Not Explanation Truth
Abstract
Autointerp simulation scoring is used as a quality signal for sparse autoencoder latent descriptions, a use that presumes the score tracks whether a description is correct. We test that presumption with a within-latent 2x2 over whether a description names the latent’s trigger token and whether its semantic gloss is true, on 16 token-driven latents of a Gemma-2-2B layer-12 GemmaScope SAE. Naming the token raises the score by 0.305, and does so for all 16 latents. Conditional on naming, swapping a true gloss for a false one changes it by 0.021 (95% CI -0.11 to +0.15). A judge-free baseline that marks tokens whose surface form appears in the description reproduces the effect. The results suggest that a high score can reflect naming the right token rather than truth of the description.
