Metraipolis › archive › dc_ff438c8171d8

Where You Measure Decides What You Measure: Position Selection in Ablation-Based SAE Evaluation — arXiv:2608.13337

Source
arXiv (en)
Original
https://arxiv.org/abs/2608.13337
License (as archived)
CC0-1.0
Archived
2026-10-19T04:27:39.046Z
SHA-256 of the archived text
22eb856715282f0d8df2c0ce6ae1d721ba1e594c90f4447bf70e425039a234e0
Stable URL of this page
https://metraipolis.com/documento/dc_ff438c8171d8

How to cite

Where You Measure Decides What You Measure: Position Selection in Ablation-Based SAE Evaluation — arXiv:2608.13337 — arXiv (en), https://arxiv.org/abs/2608.13337 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_ff438c8171d8 (Metraipolis archive dc_ff438c8171d8, sha256 22eb856715282f0d8df2c0ce6ae1d721ba1e594c90f4447bf70e425039a234e0)
BibTeX
@misc{dcff438c8171d8,
  title = {Where You Measure Decides What You Measure: Position Selection in Ablation-Based SAE Evaluation — arXiv:2608.13337},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2608.13337},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_ff438c8171d8; sha256 22eb856715282f0d8df2c0ce6ae1d721ba1e594c90f4447bf70e425039a234e0},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2608.13337. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Sparse autoencoders are meant to name the things a language model computes, and the usual way to check that a latent matters is to switch it off and see what changes. But a latent fires at many tokens, and the effect has to be measured at one of them. The convention is to measure where the latent fires hardest. That choice is almost never reported, and it is not made by the experimenter: it is made by the dictionary under evaluation. Change the dictionary and the measurement moves to a different token. We show this is not a detail. Take two sparse autoencoders released by Google for the same model and match their latents by decoder similarity: even among the pairs the two dictionaries encode almost identically, they pick different tokens for a large share of them. Two dictionaries compared under the usual protocol are therefore very often compared at different places. To separate the convention from the dictionaries we train six autoencoders from one initialisation, differing only in fitting choices, so that a latent means the same thing in each. Most of the variance such a comparison reads as "these dictionaries disagree about this latent" turns out to be the position instead: it falls from 7.6% and 11.9% of variance to near zero once every dictionary is measured at the same token. More evaluation data does not rescue it. Across a sixteenfold range of corpus sizes the dictionaries agree less about where to measure, not more, so the problem grows with scale. The correction is one line of evaluation code. We give the protocol an ablation-based causal number must report to be comparable across papers, and an audit of five published papers against it. In short: a causal number reported without its position describes the token it was taken at as much as the latent it was taken from.

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-08-14 in cs.LG).
Categorie arXiv: Machine Learning (cs.LG).
Pagina del lavoro: https://arxiv.org/abs/2608.13337 — 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 2026-10-19T04:20:35.858Z dalla pagina https://arxiv.org/catchup/cs.LG/2026-08-14?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).