Fetching from the wire…
01
02
03
04
05
06
07
08
Research2026-08-15 · source-backed
The standard protocol ablates a latent and measures effect at the token where it fires hardest, but that token is chosen by the dictionary under evaluation. Two dictionaries get compared at different places. Training six autoencoders from one initialization showed 7.6% and 11.9% of apparent inter-dictionary variance collapses to near zero once position is held fixed. Across a sixteenfold range of corpus sizes, dictionaries agreed less about where to measure. arXiv 2608.13337 The problem grows with scale, and the paper audits five published papers against a corrected protocol.
Each link below shares sources, entities, or timing with this story.
Shared entity: Training / Same source domain / Shared topic / Earlier coverage / Tension
Both cover Training; reported by the same outlet (arxiv.org); overlapping topics (paper, token).
Shared entity: Training / Same source domain / Shared topic / Earlier coverage
Both cover Training; reported by the same outlet (arxiv.org); overlapping topics (against, collaps).
Shared entity: Training / Same source domain / Earlier coverage / Tension
Both cover Training; reported by the same outlet (arxiv.org); earlier Training coverage from 2026-08-13.
Both cover Training; reported by the same outlet (arxiv.org); earlier Training coverage from 2026-08-11.
Same source domain / Shared topic / Tension
Reported by the same outlet (arxiv.org); overlapping topics (against, audit, comparison); pushes against this story (but).
Shared entity: Training / Same source domain / Earlier coverage
Both cover Training; reported by the same outlet (arxiv.org); earlier Training coverage from 2026-08-03.
Both cover Training; reported by the same outlet (arxiv.org); earlier Training coverage from 2026-07-31.
Shared entity: Sparse / Same source domain / Earlier coverage
Both cover Sparse; reported by the same outlet (arxiv.org); earlier Sparse coverage from 2026-07-27.