The Problem With Current Routing

Most AI systems today route at the wrong layer. Multi-agent pipelines route at the text level — slow, brittle, expensive. MoE architectures route at the token level — better, but still late.

There is a more fundamental level: the latent space.

What Is Latent Routing?

Latent Routing makes inference-time decisions based on the geometric position of an input’s representation in a model’s latent space.

Given input x and encoder E, the routing decision r is:

r = Router(E(x))

The Router operates on the latent vector, not raw tokens.

Three Forms

Expert Routing — extends MoE from per-token to per-concept routing.

Agent Routing — selects specialist agents based on latent-space proximity to their competence regions.

Compute Routing — exits inference early when the latent representation lands near a confident region. Same quality, less compute.

Why Now

Three trends converging: models are becoming modular, inference cost is the bottleneck, and latent space is becoming legible via mechanistic interpretability.

Latent Routing is the primitive that ties them together.


Follow @latentroute for updates.