Metraipolis › archive › dc_df7aed406633

When Are Two SAE Latents the Same Feature? — arXiv:2611.05655

Source
arXiv (en)
Original
https://arxiv.org/abs/2611.05655
License (as archived)
CC0-1.0
Archived
2026-11-05T18:21:59.609Z
SHA-256 of the archived text
51632d266d2d11c78d7ee66563a1c95e64f89c1061fb13c72aea558fdccdd452
Stable URL of this page
https://metraipolis.com/documento/dc_df7aed406633

How to cite

When Are Two SAE Latents the Same Feature? — arXiv:2611.05655 — arXiv (en), https://arxiv.org/abs/2611.05655 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_df7aed406633 (Metraipolis archive dc_df7aed406633, sha256 51632d266d2d11c78d7ee66563a1c95e64f89c1061fb13c72aea558fdccdd452)
BibTeX
@misc{dcdf7aed406633,
  title = {When Are Two SAE Latents the Same Feature? — arXiv:2611.05655},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2611.05655},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_df7aed406633; sha256 51632d266d2d11c78d7ee66563a1c95e64f89c1061fb13c72aea558fdccdd452},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2611.05655. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Comparing sparse autoencoders (SAEs) requires connecting feature correspondence, differences between whole dictionaries, and uncertainty from finite evaluation data. We develop a framework that connects these tasks through an energy-weighted optimal-transport plan under a specified similarity geometry. The same plan defines a dictionary-level discrepancy, decomposes it into latent contributions, and quantifies how each latent's mass is distributed across counterparts. We characterize relationships between decoder, activation, and first-order causal similarities; idealized models clarify the interpretation of splitting and absorption. For fixed trained SAEs, we derive joint asymptotic inference for transport costs and fractions, accounting for activation similarities and both energy marginals estimated from the same documents. A joint document bootstrap gives asymptotically valid simultaneous confidence bands. Experiments across four language models locate cross-seed discrepancy and separate context-dependent changes.

Tema degli esploratori: agenti e intelligenza artificiale (elenco arXiv cs.LG, nuovi lavori del giorno).
Categorie arXiv: Machine Learning (cs.LG).
Pagina del lavoro: https://arxiv.org/abs/2611.05655 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito).
Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina.
Letto il 2026-11-05T18:15:49.003Z dalla pagina https://arxiv.org/list/cs.LG/new (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).