Research map / Probabilistic and scientific modeling / Causal discovery

CausalTab: Pretraining Across Causal Environments for Tabular Causal Discovery

NeurIPS 2026 · Probabilistic and scientific modeling › Causal discovery · approach: Interventional causal discovery

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes TabCausal, a tabular causal discovery foundation model pretrained over broad causal environments (graph priors, mechanisms, noise, dimensions, sample sizes, observational/interventional regimes) that predicts directed edge probabilities in one forward pass, and introduces a semantic causal environment benchmark.

Problem

Existing causal discovery foundation models (CDFMs) do not consistently beat strong classical causal discovery baselines, which the authors attribute largely to narrow causal pretraining task construction.

Approach

Scale up and diversify the pretraining causal environment (graph priors, SCM mechanisms, noise, dimensions, sample sizes, intervention regimes) with dynamic task composition, and train an axial-attention transformer that maps value-mask tensors to directed edge probabilities in one pass.

Key results

• On the synthetic benchmark under mixed-interventional data, TabCausal reaches macro-averaged F1 0.952 and SHD 2.2, with average rank 1.0 versus AVICI at 2.5.
• On the semantic benchmark with mixed interventions, TabCausal achieves F1 0.971 and SHD 0.88, best among applicable methods.
• On real datasets with mixed interventions, TabCausal has F1 0.42 and SHD 40.29, best in both; observation-only F1 is 0.32, also best, while LiNGAM has the best observation-only SHD.

Surprising findings

• Under observation-only data, GIES has the highest semantic-benchmark F1 (0.698) over TabCausal (0.681).
• Several optimization-based baselines (e.g., DAGMA, SDCD) can outperform TabCausal on some high-dimensional or pfn observational settings.

Limitations

• Evidence is mostly simulator-grounded; the real-data check is scored but not a field deployment.
• Observational non-identifiability and Markov-equivalence ambiguity remain; large graphs and hard observational cases are difficult.
• Does not address hidden confounding, cyclic feedback, temporal structure, or selection effects.

Experimental setup

Datasets: CD-CSG Sachs Amazon PetShop Causal Chambers linear_gauss linear_nongauss linear_graph gp_simple gp_hard mul_noise pfn

Benchmarks: synthetic benchmark families (7 families, d in {5,10,20}) semantic causal environment benchmark real-data causal discovery datasets (CD-CSG, Sachs, PetShop, Causal Chambers) high-dimensional and sample-size scaling

Metrics: F1 SHD SID average rank

Baselines: AVICI SEA GIES PC NOTEARS NOTEARS-MLP IGSP CDIS DAGMA SDCD DCDI NoDAGS

Builds on: AVICI TabPFN NOTEARS GIES

Closest papers by meaning

• A Meta-Learning Approach to Bayesian Causal Discovery (ICLR 2025 · similarity 0.90)
• CausalRivers - Scaling up benchmarking of causal discovery for real-world time-series (ICLR 2025 · similarity 0.90)
• Integrating Background Knowledge for Scalable Causal Discovery (NeurIPS 2026 · similarity 0.90)
• Revealing Multimodal Causality with Large Language Models (NeurIPS 2025 · similarity 0.90)
• Learning Mixtures of Unknown Causal Interventions (NeurIPS 2024 · similarity 0.89)
• Learning Structural Causal Models from Ordering: Identifiable Flow Models (AAAI 2025 · similarity 0.89)