Why the answer hides in the routes

When a system runs a language model many times and has to pick one answer, it almost always votes on the text — extract the answer string from each trace, canonicalize it, and count. That works for short, canonical answers. It quietly breaks when the output is code, a proof, or anything lexically diverse: there is no clean string to vote on.

The interface, not the rule

Sparse Mixture-of-Experts models expose a second signal. To produce each token, the router picks a few experts — a record of which experts drove that token, produced online from the hidden state and never decoded to text. Hold a repeated anchor token fixed and that routing footprint still separates task, history, and reasoning effort. So the same surface token can be generated through different routes.

Majority counts answer strings; RAD aggregates routing agreement.

If routing neighborhoods line up with answer basins near the place an answer opens, you can select among completed rollouts without reading a single answer string — and you are defined exactly where string voting is not.

This is a placeholder post — replace it with your own writing. (See the RAD paper page for the full version of this idea.)