CircuitTree：将电路组合为音节转写器

AI 摘要

该研究提出 CircuitTree 数据结构与贪心、密度缩放组合算法，解决此前简单掩码并集组合电路时的密度支配与干扰问题。作者从 HuBERT 中剪枝出 39 个音素电路（15 个元音、24 个辅音），将其组合为音节分类器，实现单个电路不具备的能力。通过将 CircuitTree 串联成电路层级流水线，开放音节转写准确率达 64.8%，仅比不共享任何电路的 66.0% 上限低 1.2 个百分点，且 99.5% 的电路组件被共享。

正文节选

CircuitTree: Composing Circuits into A Syllable Transcriber 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. We propose a greedy, density-scaled composition algorithm that overcomes limitations of previous attempts to compose circuits by naive union, then generalize i

发布时间：2026-10-12 12:00

抓取时间：2026-10-12 12:13

来源机构：arXiv