Search papers, labs, and topics across Lattice.
This study rigorously evaluates the performance of attention-only transformer architectures, specifically Simple Attention Networks (SANs), by controlling for parameters, compute, and depth against standard transformers. The findings reveal that while SANs initially lag behind standard transformers in performance, reallocating resources from feed-forward layers to attention depth significantly narrows the performance gap to a mere 0.006 nats. The research highlights that the remaining performance deficit is primarily due to parametric recall issues, indicating that attention mechanisms excel in context-grounded tasks but struggle with knowledge retrieval from weights.
Attention-only transformers can match standard architectures in performance when resource allocation favors attention depth, challenging assumptions about the necessity of feed-forward layers.
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.
