bibkey: “geiger2025causal” authors: “Atticus Geiger; Duligur Ibeling; Amir Zur; Maheep Chaudhary; Sonakshi Chauhan; Jing Huang; Aryaman Arora; Zhengxuan Wu; Noah Goodman; Christopher Potts; Thomas Icard” year: 2025 title: “Causal Abstraction: A Theoretical Foundation for Mechanistic Interpretability” doi: null claim: “Causal abstraction transports solution sets under allowed interventions through state and intervention maps.” strata_touched: [] license: “citation-only” triage: “anchor” url: “https://www.jmlr.org/papers/v26/23-0058.html”

Causal Abstraction: A Theoretical Foundation for Mechanistic Interpretability

Verified locator

Journal of Machine Learning Research 26(83), 1–64, published May 2025. Definition 25, page 13, specifies partial surjective state and intervention maps, preservation of intervention order, and equality of transported solution sets. Definition 33, page 17, defines constructive abstraction. Definition 48, page 23, defines interchange intervention accuracy as an expectation under a chosen intervention distribution; this average is not an unrestricted full-domain guarantee.