AI/ ai · machine-learning · transformers · research

Linear Attention Gets a Checkup - and a Modest Upgrade

A new study benchmarks four recurrent linear-attention architectures and finds a routing trick that quietly lowers validation loss.

Researchers have published a head-to-head comparison of softmax attention and four linear-attention alternatives, plus a small routing trick that ekes out measurable gains.

The paper pits DeltaNet, Gated DeltaNet, Kimi Delta Attention, and Gated DeltaNet-2 against each other using a shared notation that makes their differences legible. Tests ran on 350M-parameter models trained for 15 billion tokens, with additional runs at 1.3B and 3B parameters for DeltaNet. The headline results: Kimi Delta Attention paired with the Muon optimizer reached the lowest validation loss in the main sweep, while a pure Gated DeltaNet stack trained with AdamW posted the highest training throughput. Hybrid stacks — mixing linear and standard attention layers — generally improved loss but at a throughput cost.

The throughput angle matters because the entire pitch for linear attention is escaping the quadratic cost of standard self-attention at long context. Any architecture that trades that speed advantage back for better loss numbers needs a compelling reason to exist. The paper also introduces Cross-Layer Value Routing, which forwards a write value between layers rather than a write error — a small change that lowers validation loss for both DeltaNet and Gated DeltaNet without the overhead of a full hybrid stack.

The authors are careful to note they did not run an inference-speed benchmark, only training throughput — a meaningful caveat when the community is as interested in fast serving as in fast training.

TR

The Revision

Written by an AI system from the public sources credited above. How we write →