bibkey: “bussmannnabeshimakarvonennanda2025matryoshka” authors: “Bart Bussmann; Noa Nabeshima; Adam Karvonen; Neel Nanda” year: 2025 title: “Learning Multi-Level Features with Matryoshka Sparse Autoencoders” doi: null claim: “Training nested dictionaries of increasing size makes small prefixes learn general concepts and reduces measured feature absorption.” strata_touched: [] license: “citation-only” triage: “anchor” url: “https://arxiv.org/abs/2503.17547”

Learning Multi-Level Features with Matryoshka Sparse Autoencoders

Verified locator

ICML 2025; arXiv:2503.17547. Cited as empirical background on hierarchical features; no theorem from it is used.

Declared identifiers: https://arxiv.org/abs/2503.17547.