Learning causal abstractions from data

Develop methods for learning causal abstractions, including the partition and corresponding coarse causal graph, directly from data without assuming the partition in advance or imposing strong parametric assumptions.

Background

The paper situates its contribution within a literature in which causal abstractions and cluster DAGs are primarily studied theoretically or under a known partition. Existing methods either take the partition as input, fix it to preserve a chosen estimand, or learn abstraction maps under strong parametric assumptions. Consequently, the broader problem of discovering an abstraction from data remains unresolved, motivating the score-based method developed in the paper for a particular interventional, linear-Gaussian setting.

References

Most of this work is theoretical in spirit, and learning an abstraction from data remains largely open: existing methods either take the partition as input \citep{entner_estimating_2012,anand_causal_2023,wahl_vector_2023}, fix it to preserve a chosen estimand \citep{tikka_clustering_2023}, or learn abstraction maps under strong parametric assumptions \citep{niu_learning_2022,massidda_learning_2024,xia_neural_2024,zennaro_jointly_2023,felice_causally_2025,dacunto_causal_2025,bing_structural_2026,felekis_distributionally_2025}.

— Score-Based Learning of Cluster DAGs from Interventions  (2610.11947 - Tedesco et al., 8 Oct 2026) in Section 1, Introduction

The main open problems are to recover the blocks and $(\tau,\omega)$ under explicit identifiability assumptions; extend the maximality result to stochastic SCMs, soft interventions, and quantitatively approximate abstractions; and evaluate concept-directed steering and transport in real neural representations, where exact generator comparisons are unavailable.

— When Do Learned Blocks Preserve Causal Meaning? Block Identifiability and Constructive Causal Abstractions  (2610.18155 - Zaghw et al., 13 Oct 2026) in Discussion, paragraph “Future work”