bibkey: “leask2025canonical” authors: “Patrick Leask; Bart Bussmann; Michael Pearce; Joseph Bloom; Curt Tigges; Noura Al Moubayed; Lee Sharkey; Neel Nanda” year: 2025 title: “Sparse Autoencoders Do Not Find Canonical Units of Analysis” doi: null claim: “SAE stitching and meta-SAEs show that dictionaries of different widths are neither complete nor atomic, so sparse autoencoders do not yield a canonical feature inventory.” strata_touched: [] license: “citation-only” triage: “anchor” url: “https://arxiv.org/abs/2502.04878”

Sparse Autoencoders Do Not Find Canonical Units of Analysis

Verified locator

ICLR 2025; arXiv:2502.04878. Cited as empirical background on non-canonical units; not used as a mathematical non-identifiability theorem.

Declared identifiers: https://arxiv.org/abs/2502.04878.