The goal is to decompose a monolithic neural network into interpretable subnetworks, each performing an identifiable subtask.
We introduce a practical composition method that can merge an arbitrary number of circuits into a single model. The core is a greedy, density‑scaled composition algorithm that overcomes the limitations of naive union. It is generalized to N circuits via the CircuitTree data structure, which follows a $\Theta(N\log N)$ tree‑compose schedule.
In experiments we prune 39 phoneme‑level circuits from HuBERT and use CircuitTree to compose them into a syllable classifier—a capability none of the individual circuits possesses.
Chaining multiple CircuitTrees into a hierarchy‑of‑circuits pipeline yields 64.8% accuracy on open‑syllable transcription, only 1.2 points below a 66.0% ceiling that shares no circuitry, while 99.5% of circuit components are shared.
Review: This work demonstrates the promise of modular circuit composition for speech tasks and offers a fresh avenue for building interpretable models.