cd /news/artificial-intelligence/a-controlled-study-of-attention-only… · home topics artificial-intelligence article
[ARTICLE · art-67941] src=arxiv.org ↗ pub= topic=artificial-intelligence verified=true sentiment=· neutral

A Controlled Study of Attention-Only Transformers

A controlled study pretraining attention-only decoder transformers (Simple Attention Networks) against standard transformers matched for parameters, compute, and depth found that reallocating the feed-forward budget into attention depth closes the performance gap to 0.006 nats (0.27% of loss) at matched parameters, reproducible to one part in ten thousand. The remaining deficit is localized to parametric recall, with attention-only models better on context-grounded answers and worse on knowledge-dense web text, where a pre-registered test on fineweb-edu measured a 0.040 nat gap. The study concludes that within the tested regime up to 105B tokens and 87M parameters, attention does the rest.

read2 min views1 publishedJul 22, 2026
A Controlled Study of Attention-Only Transformers
Image: source
[Submitted on 20 Jul 2026]


[View PDF](/pdf/2607.18363)

[HTML (experimental)](https://arxiv.org/html/2607.18363v1)

Abstract:Feed-forward networks hold two thirds of a transformer's non-embedding parameters, yet the architecture has not received a necessity test that controls parameters, compute, and depth at once. We pretrain attention-only decoder transformers (Simple Attention Networks, SANs) against standard transformers matched separately for parameter count, training FLOPs, and depth (2 to 48 layers), for up to 105B tokens at 6M to 87M parameters. Deleting feed-forward layers in place is costly: the standard transformer leads by 0.47 nats at matched depth and 0.26 nats at matched FLOPs. Reallocating the freed budget into attention depth closes the gap: at matched parameters the difference is 0.006 nats (0.27 percent of loss), reproducible to one part in ten thousand across seed pairs, shrinking across 5B, 30B, and 105B budgets, and holding near 0.02 nats across a 29x size range. Three measurements localize the remaining gap to parametric recall: attention-only models are better on context-grounded answers and worse where knowledge must come from weights. Weight spectra show why: routing matrices (Q/K) crystallize early, content matrices accumulate rank slowly, and removing feed-forward layers relocates this accumulation to the attention output projection. QK-normalization, not feed-forward layers or residual gating, keeps 48-layer attention-only stacks trainable. The deficit concentrates on low-context query prediction and localizes there entirely by the largest budget. A pre-registered test confirms the account: it predicts a 0.02 to 0.05 nat gap on knowledge-dense web text; a matched pair trained on fineweb-edu measures 0.040. Within the tested regime, attention does the rest.

Current browse context:

cs.LG

References & Citations

...

Bibliographic Explorer

(What is the Explorer?) Connected Papers

(What is Connected Papers?) Litmaps

(What is Litmaps?) scite Smart Citations

(What are Smart Citations?)# Code, Data and Media Associated with this Article alphaXiv

(What is alphaXiv?) CatalyzeX Code Finder for Papers

(What is CatalyzeX?) DagsHub

(What is DagsHub?) Gotit.pub

(What is GotitPub?) Hugging Face

(What is Huggingface?) ScienceCast

(What is ScienceCast?)# Demos Influence Flower

(What are Influence Flowers?) CORE Recommender

(What is CORE?) IArxiv Recommender

(What is IArxiv?)# arXivLabs: experimental projects with community collaborators arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.

── more in #artificial-intelligence 4 stories · sorted by recency
── more on @simple attention networks 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/a-controlled-study-o…] indexed:0 read:2min 2026-07-22 ·