bibkey: “chanin2024toymodels” authors: “David Chanin” year: 2024 title: “Toy Models of Feature Absorption in SAEs” doi: null claim: “In toy models with controlled co-occurrence, absorption appears when features co-occur, and partial absorption appears when co-occurring feature magnitudes vary.” strata_touched: [] license: “citation-only” triage: “anchor” url: “https://www.lesswrong.com/posts/kcg58WhRxFA9hv9vN/toy-models-of-feature-absorption-in-saes”

Toy Models of Feature Absorption in SAEs

Verified locator

LessWrong, 7 October 2024. Cited for the empirical report of absorption and partial absorption under co-occurrence and variable magnitudes; used only as a boundary example in the fiber-law volume.

Declared identifiers: https://www.lesswrong.com/posts/kcg58WhRxFA9hv9vN/toy-models-of-feature-absorption-in-saes.