Relational Structural Causal Models, Ejaz & Bareinboim 2026

Relational Structural Causal Models, a paper by Adiba Ejaz and Elias Bareinboim submitted to arXiv on 12 Jun 2026 (arXiv:2606.14892), extends Pearl-style structural causal models to settings where objects and their relations vary. The authors formalize when a causal model can both support interventions and counterfactuals and generalize to unseen combinations of objects, then propose a learned neural approach and evaluate it on simulated traffic scenes.

What the paper does

The paper develops a formal notion of relational structural causal models, an extension of structural causal models to relational domains. It shows that answers to not only causal but also observational queries about unseen combinations of objects cannot be identified without further assumptions. To enable identification, including when unobserved confounding is present, the authors define relational causal graphs and derive symbolic identification criteria.

The work also proposes relational neural causal models, which the authors describe as a provably correct approach. The paper reports that these relational neural causal models outperform non-relational baselines on simulated traffic scenes containing varying cars, signals, and pedestrians. The submission lists the paper as part of the Proceedings of the Forty-Third International Conference on Machine Learning and provides an arXiv DOI link (https://doi.org/10.48550/arXiv.2606.14892).

Technical contributions

The core theoretical contribution is a set of symbolic identification criteria for relational settings. Those criteria operate on relational causal graphs, a formalism the paper defines to represent objects and the relations among them when causal queries involve unseen combinations. The authors place emphasis on identification both for observational queries and for causal queries under interventions, and they explicitly account for the possibility of unobserved confounding.

On the empirical side, the paper introduces relational neural causal models as an algorithmic instantiation of the theory. The evaluation uses simulated traffic scenes that vary the number and arrangement of cars, traffic signals, and pedestrians. According to the paper, relational neural causal models are provably correct within the formal framework and achieve superior performance compared with non-relational baselines on those simulations.

Why it matters

Causal models that generalize across combinations of objects are essential where environments are compositional, such as traffic, social networks, or structured scientific data. By tying symbolic identification criteria to a neural implementation, the paper bridges a gap between causal theory and practical learning in relational domains. If the relational causal graphs and the associated identification rules hold up in broader settings, they provide a principled path to answer counterfactual and intervention queries on scenes the model has not seen before.

What to watch

Look for follow-up work testing relational neural causal models on real-world relational data beyond simulated traffic scenes, and for independent replication of the symbolic identification rules. Another signal will be whether the definitions of relational causal graphs are adopted in subsequent causal-inference or structured-learning papers presented at the Forty-Third International Conference on Machine Learning.

References and provenance: the paper is titled "Relational Structural Causal Models," authored by Adiba Ejaz and Elias Bareinboim, submitted to arXiv on 12 Jun 2026 as arXiv:2606.14892 and linked to the Proceedings of the Forty-Third International Conference on Machine Learning.