Sparsely gated tiny linear experts
Researchers propose sparsely gated linear neurons (sgatlin) as a radical departure from dense mixture-of-experts architectures, replacing nonlinear expert layers with single-neuron linear gates. The approach trades conventional neural network design for extreme sparsity, achieving better perplexity-per-compute than standard transformer feedforward layers across multiple scales. The counterintuitive finding that removing nonlinearity improves both efficiency and interpretability challenges assumptions about what makes expert routing effective, potentially reshaping how practitioners scale language models under compute constraints.
Modelwire context
ExplainerThe key buried detail is that sgatlin's interpretability gains are a direct structural consequence of linearity, not a post-hoc analysis feature bolted on. Because each expert is a single neuron with a linear gate, the routing decision is inspectable by construction, which is a fundamentally different claim than MoE systems that add interpretability tooling on top of opaque expert networks.
The MoE space in our recent coverage has been moving in two directions simultaneously: domain-specific routing (as in GC-MoE from early June, which adds genomic priors to steer expert selection) and continual learning stability (SETA's sparse expert decomposition for preventing catastrophic forgetting). Sgatlin cuts against both trends by questioning whether expert complexity is the right axis to optimize at all. Where those papers add structure to routing, this paper strips it out. That tension is worth tracking because the efficiency argument here, better perplexity-per-compute at multiple scales, is exactly the metric that drives infrastructure decisions in production deployments.
Watch whether any of the major scaling labs reproduce the perplexity-per-compute advantage on models above 7B parameters in the next six months. If the gains compress or disappear at larger scales, the result is likely a compute-regime artifact rather than a general architectural principle.
Coverage we drew on
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
Mentionssgatlin · mixture of experts · transformer
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.