Chiosa
ML · learning path

Match routing decisions to the evidence available at each step

7 min3 papersbeginner
Published on Chiosa Updated

An agent may need to choose a model for a home-control step, decide how to ask owners of separate tables for an answer, or send a question to a larger model. Each choice can consume additional computation, yet each exposes different evidence about where that work may help.

A useful first answer is to match the routing decision to the variation the system can observe: a known step, a dependency between owners, or a question-level confidence signal.

These studies test those choices in different settings, so their outcomes guide conditional decisions rather than a shared performance ranking.

What you will understand

  • How can a known call site or the ownership of needed tables shape a routing choice?

  • What evidence shows whether returned rows or low confidence help identify useful additional work?

  • Which limits should accompany a routing result before applying it to another deployment or task?

Locate the choice before choosing a route

Imagine a home-control agent receiving a request to turn on a device. Its system already has separate steps for recognizing the request, selecting an action, planning a longer task, and writing code for that plan.

One deployment decision is whether every step needs the same hosted model. Another is whether a smaller local model can handle a particular step while a hosted model serves the steps that ask it to produce a longer plan or code. The system knows which step it is about to run.

Now imagine an organization whose people hold different tables. A coordinator receives a question that needs rows from two owners, but its public catalog lists only table names and columns, not the rows.

It can ask owners in parallel and combine what comes back, or ask them one after another. If a row from the first owner reveals a value needed to find a matching row at the second, the order of those questions matters. The relevant variation lies in the requested data.

A third choice arises when every multiple-choice question first goes to a small model. The system could accept its answer or ask a larger model from the same family to answer that question.

Here the choice changes from one question to the next, and the small model's apparent uncertainty is a possible clue. Across these examples, the shared problem is where to make a routing decision and what evidence justifies spending additional computation there.

Where the lesson goes next

  • Use a signal that fits the decision

  • Read the evidence at its measured scale

  • Keep the decision within its tested setting