CircuitTree: Composing Circuits into A Syllable Transcriber

Computer Science > Machine Learning arXiv:2610.12520 (cs) [Submitted on 7 Oct 2026] Title: CircuitTree: Composing Circuits into A Syllable Transcriber Authors: Jiankun Wei, Ewan Dunbar, Gerald Penn.

计算机科学 > 机器学习 arXiv:2610.12520 (cs) [提交于 2026 年 10 月 7 日] 标题：CircuitTree：将电路组合成音节转录器 作者：Jiankun Wei, Ewan Dunbar, Gerald Penn。

Abstract: The goal of circuit discovery is to decompose a monolithic neural network into tiny modular subnetworks which are interpretable in the sense that each does an identifiable subtask. We develop a practical method for composing arbitrarily many circuits into a single model.

摘要：电路发现的目标是将单体神经网络分解为微小的模块化子网络，这些子网络具有可解释性，即每个子网络都能执行可识别的子任务。我们开发了一种将任意数量的电路组合成单一模型的实用方法。

We propose a greedy, density-scaled composition algorithm that overcomes limitations of previous attempts to compose circuits by naive union, then generalize it to $N$ circuits via CircuitTree, a new data structure built on a $\Theta(N\log N)$ tree-compose schedule.

我们提出了一种贪婪的、密度缩放的组合算法，克服了以往通过简单并集组合电路的局限性，并通过 CircuitTree 将其推广到 $N$ 个电路，这是一种基于 $\Theta(N\log N)$ 树组合调度构建的新型数据结构。

We compose circuits pruned from HuBERT for classifying individual phonemes into a syllable classifier, a capability none of the 39 phoneme circuits has individually. Chaining CircuitTrees into a hierarchy-of-circuits pipeline, we transcribe open syllables at 64.8% accuracy, within 1.2 points of a 66.0% ceiling that shares no circuitry at all, with 99.5% of circuit components shared.

我们将从 HuBERT 中剪枝得到的用于分类单个音素的电路组合成一个音节分类器，这是 39 个音素电路中没有任何一个单独具备的能力。通过将 CircuitTrees 链接成电路层级流水线，我们以 64.8% 的准确率转录开音节，这与完全不共享电路的 66.0% 的上限仅差 1.2 个百分点，同时共享了 99.5% 的电路组件。