arXiv：cs.LG· Jiankun Wei, Ewan Dunbar, Gerald Penn·· 11 小时前
CircuitTree：将电路组合成音节转写器
CircuitTree: Composing Circuits into A Syllable Transcriber
AI 导读
研究者提出 CircuitTree，一种基于 Θ(N log N) 树组合调度的数据结构，可将任意多个电路组合进单一模型，并用贪心、密度缩放的组合算法克服此前简单并集组合电路的局限。
正文
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 it to $N$ circuits via CircuitTree, a new data structure built on a $\Theta(N\log N)$ tree-compose schedule. 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.
| Subjects: | Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Sound (cs.SD) |
| Cite as: | arXiv:2610.12520 [cs.LG] |
| (or arXiv:2610.12520v1 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2610.12520 arXiv-issued DOI via DataCite |
Submission history
From: Jiankun Wei [view email]
[v1]
Wed, 7 Oct 2026 21:29:44 UTC (148 KB)
来源：arXiv：cs.LG · arxiv.org