All Circuits Lead to Rome: Rethinking Functional Anisotropy in Circuit and Sheaf Discovery for LLMs

Xi Chen, Mingyu Jin, Jingcheng Niu, Yutong Yin, Jinman Zhao, Bangwei Guo, Dimitris N. Metaxas, Zhaoran Wang, Yutao Yue, Gerald Penn · 2026

Read All Circuits Lead to Rome: Rethinking Functional Anisotropy in Circuit and Sheaf Discovery for LLMs by Xi Chen, Mingyu Jin, Jingcheng Niu, Yutong Yin, Jinman Zhao, Bangwei Guo, Dimitris N. Metaxas, Zhaoran Wang, Yutao Yue, Gerald Penn (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-12671.txt.

Introduction

Chen and nine colleagues challenge the Functional Anisotropy Hypothesis in circuit hunt for language models. That idea holds that functions sit in a unique or near-unique internal mechanism. The team shows one task can instead be backed by multiple distinct circuits that are at once faithful, sparse and complete. Their Overlap-Aware Sheaf Repulsion penalises overlap across runs, uncovering competing paths with strong task scores but minimal shared structure across common tests.

Catalog record

CREATORXi Chen, Mingyu Jin, Jingcheng Niu, Yutong Yin, Jinman Zhao, Bangwei Guo, Dimitris N. Metaxas, Zhaoran Wang, Yutao Yue, Gerald Penn
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE126831
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.12671

Browse the full library · About the Tower of Babel