← 研踪 PaperTrailarXiv:2608.22007 · v3
The Communication Map of a Transformer
首次提交 2026-08-22 23:24
版本更新 2026-11-07 22:52
雷达观测 2026-11-10 11:14
摘要
The components of a transformer communicate by writing to and reading from a shared residual stream, and the mechanistic interpretability literature has mapped these connections by hand, one circuit at a time. We present the communication map, which charts every potential communication channel from the geometry of the model's weights alone. Its unit of analysis is the whole transformer rather than the individual head or circuit. Every ordered writer-reader pair, across all 18 connection classes from head-to-head to neuron-to-neuron, receives a coupling coefficient that generalizes the composition score of Elhage et al. (2021), with an exact chance level. The census finds that 70-89% of head pairs are oriented far from chance, some coupled strongly and others actively avoiding each other. The map recovers the known induction circuits blind from the strongest head-to-head couplings and groups the heads into communities, and ablating one such community destroys the model's in-context copying. As an application, we pool the coupling coefficients of every head to identify a two-dimensional residual stream subspace, distinct from those found by activation PCA or outlier dimensions, whose deletion abolishes induction in six models up to Pythia-6.9B. We release the map, the statistical machinery, and the intervention suite.
作者备注：33 pages (12 pages of main text). Code and results: https://github.com/richardzhewang/communication-map
观测到的版本
相邻记录可能跳过中间版本；首次发现 v2 及以上也会标为版本更新。
v3 · 2026-11-07 22:52 · 最新观测版本
The Communication Map of a Transformer
The components of a transformer communicate by writing to and reading from a shared residual stream, and the mechanistic interpretability literature has mapped these connections by hand, one circuit at a time. We present the communication map, which charts every potential communication channel from the geometry of the model's weights alone. Its unit of analysis is the whole transformer rather than the individual head or circuit. Every ordered writer-reader pair, across all 18 connection classes from head-to-head to neuron-to-neuron, receives a coupling coefficient that generalizes the composition score of Elhage et al. (2021), with an exact chance level. The census finds that 70-89% of head pairs are oriented far from chance, some coupled strongly and others actively avoiding each other. The map recovers the known induction circuits blind from the strongest head-to-head couplings and groups the heads into communities, and ablating one such community destroys the model's in-context copying. As an application, we pool the coupling coefficients of every head to identify a two-dimensional residual stream subspace, distinct from those found by activation PCA or outlier dimensions, whose deletion abolishes induction in six models up to Pythia-6.9B. We release the map, the statistical machinery, and the intervention suite.
查看此版本解读 →查看此版本原文 ↗