MURANO: Design, Run, and Reproduce Mechanistic Interpretability Experiments as Composable Pipelines
2026-08-31 • Computation and Language
Computation and Language
AI summaryⓘ
The authors introduce Murano, an open-source tool that helps researchers study and understand how large language models work step-by-step. Unlike other tools that focus on only parts of the process, Murano combines loading data, tracking what happens inside the model, figuring out important parts, making changes, and checking results all in one system. It organizes these tasks as connected steps that share information smoothly, making it easier to run and reproduce complex experiments. The authors show how Murano works by repeating two known studies and running an example with a simple model.
mechanistic interpretabilitylarge language modelsattributioninterventionpipelineautoencoderopen-source frameworkreproducibilitymachine learning libraries
Authors
Alireza Bayat Makou, Emirhan Böge, Phu Gia Hoang, Federico Tiblias, Jingcheng Niu, Subhabrata Dutta, Richard Eckart de Castilho, Iryna Gurevych
Abstract
This paper presents Murano, an open source framework for designing, running, and reproducing mechanistic interpretability studies of large language models, intended for researchers across disciplines. These studies often combine loading, recording, attribution, intervention, and evaluation, while existing libraries tend to focus on different parts of this workflow. As a result, researchers using several libraries may need to adapt outputs from one for use by another. To bridge this gap, Murano represents operations from these five areas as composable steps. Steps exchange named result artifacts and declare the inputs they require and the outputs they produce. A pipeline executes its steps in the order supplied, and Murano uses canonical addresses when component identities pass between operations. Murano builds on existing interpretability and machine learning libraries. We demonstrate Murano through two reproductions of established interpretability studies and one illustrative sparse autoencoder case study.