CausaLab: A Scalable Environment for Interactive Causal Discovery Toward AI Scientists

Researchers introduced CausaLab, a scalable synthetic benchmark for interactive causal discovery by LLM agents.

Agents must use observational data and interventions to predict outcomes, reconstruct the causal graph, and infer the equations behind hidden structural causal models.

Results show strong models can make accurate predictions without truly recovering the causal mechanism.

Mixed observation-and-intervention strategies, plus consistency checks, improve structural fidelity and expose a gap in current AI reasoning.