Use What You Know: Causal Foundation Models with Partial Graphs

Researchers have introduced a method to condition Causal Foundation Models on domain knowledge, enabling a single general-purpose model to match the performance of specialized causal estimators, according to a paper revised in June 2026 [1][2]. Causal Foundation Models, or CFMs, aim to unify causal discovery and inference by amortising both tasks into a single step [2]. Until now, these models lacked a mechanism to incorporate existing domain expertise, which could result in suboptimal predictions [1][2]. Arik Reuter and collaborators detail an approach that bridges this gap by conditioning CFMs on causal information such as a causal graph or ancestral relationships [1][2]. The work was first submitted to arXiv on 16 February 2026 and updated on 25 June 2026 [1]. The researchers systematically evaluated several conditioning strategies [2]. They found that injecting learnable biases into the attention mechanism, combined with a graph-convolutional encoder, proved highly effective for utilizing both full and partial causal information [2]. This design allows the model to leverage whatever domain knowledge is available, even when a complete causal graph is not accessible [1][2]. The broader field of artificial intelligence has long pursued systems capable of reasoning and decision-making under uncertainty [3]. Causal inference, which moves beyond correlation to ask what would happen under an intervention, represents a distinct challenge from the pattern recognition tasks that drove the deep learning boom after 2012 [3]. The scientific method itself relies on hypothesis testing and experimental validation to establish causal relationships, a process that the new conditioning technique seeks to approximate in a data-driven manner [5]. In their experiments, the conditioned CFM matched the performance of specialized models that were trained on specific causal structures [1][2]. The authors state that the work addresses a central hurdle on the path toward all-in-one causal foundation models: the capability to answer causal queries while effectively leveraging any amount of domain expertise [2]. The paper was released through arXivLabs, a framework for community collaborators to develop features on the preprint platform [1].