Skip to content

Reproductions

The Reproduction Gallery collects end-to-end reproductions of published mechanistic interpretability results, implemented with Murano pipelines. Each entry in the gallery documents the paper, the experimental setup, and the results obtained, and comes with a runnable notebook.

Browse the gallery

Interpretability moves fast, and papers rarely agree on APIs, data formats, or evaluation protocols. The gallery exists so that anyone can:

  • Run published experiments in a few lines — Murano’s unified Pipeline API collapses activation recording, probing, steering, and evaluation into declarative composable steps.
  • Compare results across papers — same dataset format, same model loader, same metrics.
  • Extend and remix — every reproduction ships as a notebook you can fork and modify.

Each gallery entry contains:

SectionWhat it covers
Paper metadataVenue, authors, BibTeX, links to arXiv / OpenReview / original code
TL;DRA 2–3 sentence summary of the core claim
AbstractVerbatim paper abstract for context
Reproducing with MuranoStep-by-step walkthrough with runnable code
Key resultsA table comparing paper vs. reproduction numbers
CitationBibTeX entry for the original work

If you’ve reproduced an interpretability paper with Murano — or you’re using Murano in your own research and want to share a reference implementation — we’d love to feature it.

Submit a reproduction →