Guardado en:
Detalles Bibliográficos
Autores principales: Gonzalez, ML Nissen, Albuquerque, Melwina, Wroe, Laurence, Cohen, Jacob Meyer, Smith, Logan Riggs, Dooms, Thomas
Formato: Preprint
Publicado: 2026
Materias:
Acceso en línea:https://arxiv.org/abs/2605.15183
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866914567706116096
author Gonzalez, ML Nissen
Albuquerque, Melwina
Wroe, Laurence
Cohen, Jacob Meyer
Smith, Logan Riggs
Dooms, Thomas
author_facet Gonzalez, ML Nissen
Albuquerque, Melwina
Wroe, Laurence
Cohen, Jacob Meyer
Smith, Logan Riggs
Dooms, Thomas
contents Mechanistic interpretability aims to break models into meaningful parts; verifying that two such parts implement the same computation is a prerequisite. Existing similarity measures evaluate either empirical behaviour, leaving them blind to out-of-distribution mechanisms, or basis-dependent parameters, meaning they disregard weight-space symmetries. To address these issues for the class of tensor-based models, we introduce a weight-based metric, tensor similarity, that is invariant to such symmetries. This metric captures global functional equivalence and accounts for cross-layer mechanisms using an efficient recursive algorithm. Empirically, tensor similarity tracks functional training dynamics, such as grokking and backdoor insertion, with higher fidelity than existing metrics. This reduces measuring similarity and verifying faithfulness into a solved algebraic problem rather than one of empirical approximation.
format Preprint
id arxiv_https___arxiv_org_abs_2605_15183
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle When Are Two Networks the Same? Tensor Similarity for Mechanistic Interpretability
Gonzalez, ML Nissen
Albuquerque, Melwina
Wroe, Laurence
Cohen, Jacob Meyer
Smith, Logan Riggs
Dooms, Thomas
Machine Learning
Mechanistic interpretability aims to break models into meaningful parts; verifying that two such parts implement the same computation is a prerequisite. Existing similarity measures evaluate either empirical behaviour, leaving them blind to out-of-distribution mechanisms, or basis-dependent parameters, meaning they disregard weight-space symmetries. To address these issues for the class of tensor-based models, we introduce a weight-based metric, tensor similarity, that is invariant to such symmetries. This metric captures global functional equivalence and accounts for cross-layer mechanisms using an efficient recursive algorithm. Empirically, tensor similarity tracks functional training dynamics, such as grokking and backdoor insertion, with higher fidelity than existing metrics. This reduces measuring similarity and verifying faithfulness into a solved algebraic problem rather than one of empirical approximation.
title When Are Two Networks the Same? Tensor Similarity for Mechanistic Interpretability
topic Machine Learning
url https://arxiv.org/abs/2605.15183