Reproduce “When Do Learned Blocks Preserve Causal Meaning? Block Identifiability and Constructive Causal Abstractions” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Karim Zaghw, Wieland Brendel, Patrik Reizinger

The paper characterizes when block-level causal abstractions survive representational changes. Its theorem shows that invertible transformations within blocks and compatible coordinate permutations preserve every abstraction on fixed blocks when alignment and interventions are transported. In a synthetic transfer test, recipient seed 2 yields intervention NMSE 0.011931, versus 0.007407 for reference seed 0; this is approximate evidence, not a general real-world method.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.