Metraipolis › archive › dc_aa77ae217324

Which Question Is Your Attention Metric Answering? Attention Rows as Compositional Data — arXiv:2608.14712

Source
arXiv (en)
Original
https://arxiv.org/abs/2608.14712
License (as archived)
CC0-1.0
Archived
2026-10-18T17:52:12.623Z
SHA-256 of the archived text
1c2bbb96856ef81fe322861fb39d79c705e9784a50acce46b58770aa026bc33d
Stable URL of this page
https://metraipolis.com/documento/dc_aa77ae217324

How to cite

Which Question Is Your Attention Metric Answering? Attention Rows as Compositional Data — arXiv:2608.14712 — arXiv (en), https://arxiv.org/abs/2608.14712 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_aa77ae217324 (Metraipolis archive dc_aa77ae217324, sha256 1c2bbb96856ef81fe322861fb39d79c705e9784a50acce46b58770aa026bc33d)
BibTeX
@misc{dcaa77ae217324,
  title = {Which Question Is Your Attention Metric Answering? Attention Rows as Compositional Data — arXiv:2608.14712},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2608.14712},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_aa77ae217324; sha256 1c2bbb96856ef81fe322861fb39d79c705e9784a50acce46b58770aa026bc33d},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2608.14712. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Each row of a transformer's attention matrix is a probability distribution over tokens, and in trained models most of that probability lands on a single \emph{sink} token, usually the first. Standard tools for comparing attention rows (cosine similarity, Jensen--Shannon divergence, Shannon entropy) therefore hinge on a choice papers rarely report: keep the sink, or drop it and renormalize. This choice can reverse conclusions. On ten pretrained models from five families, 17--47% of verdicts about which of two heads is more similar flip with the convention, and the most prominent structure in a standard BERT head-clustering pipeline is an artifact of it. The reason is that one-number summaries mix two questions: how much attention the sink takes, and how the rest is divided among the content tokens. Treating rows as compositional data separates them exactly: the Aitchison distance splits orthogonally into a sink term and a content term, entropy splits by an exact identity, and the content distance is characterized by invariances the transformer itself possesses. The separation matters in practice: most measured entropy collapse during training is the sink growing, not attention sharpening (30% of the drop at 70M parameters, 95% at 1B, 79% at 1.4B), and pruning heads with the wrong channel can inflate perplexity more than a hundredfold. We map where each convention is safe, test a frozen out-of-sample predictor (one confirmation, one abstention, one failure), and release code regenerating every number.

Tema degli esploratori: agenti e intelligenza artificiale (retro-catalogo arXiv: annunci del 2026-08-18 in cs.AI).
Categorie arXiv: Computation and Language (cs.CL) ; Artificial Intelligence (cs.AI); Machine Learning (cs.LG); Statistics Theory (math.ST).
Pagina del lavoro: https://arxiv.org/abs/2608.14712 — 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 2026-10-18T16:29:13.946Z dalla pagina https://arxiv.org/catchup/cs.AI/2026-08-18?abs=True (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).