Expert routing
Extends MoE from per-token switching to per-concept specialist selection.
Hey there, meet A.R.I.A.
Mainframe's latent routing interface agent
Inference decisions should happen where meaning forms. Route through latent space.
Latent Space
Expert Routing
Agent Routing
Compute Routing
Early Exit
Inference
Latent Space
Expert Routing
Agent Routing
Compute Routing
Early Exit
Inference
What is latent routing?
Route the representation, not the sentence.
Most AI systems route too late. Multi-agent pipelines route at the text layer, which can be slow and brittle. MoE systems route per token, which is better, but still misses the level where concepts are already shaped.
Given input x and encoder E
r = Router(E(x))
The router operates on the latent vector, not raw tokens.
Three forms
One primitive, three routing moves.
Agent routing
Selects specialist agents by proximity to their latent competence regions.
Compute routing
Exits inference early when the representation lands near a confident region.
Why now
Modularity, cost pressure, and legible latent space.
Models are becoming modular, inference cost is the bottleneck, and mechanistic interpretability is making representation geometry easier to inspect.
Opening
Input becomes geometry. Geometry becomes a decision.
Available now
Latent routing brief
A compact visual explainer for inference-time routing through representation space, inspired by the article "What Is Latent Routing?".
Return to top