CausalArena: Benchmarking Causal Discovery in the Foundation Model Era

As causal discovery foundation models emerge, static benchmarks struggle to separate genuine structural reasoning from pretraining overlap. CausalArena introduces an evolving evaluation suite spanning synthetic graphs, semantic operational environments, formula-grounded scientific mechanisms, and real-world datasets. Experiments across classical, neural, and pretrained methods show substantial ranking shifts between regimes, highlighting the fragility of performance claims and the necessity of multi-environment testing.