CasualSynth: Generating Structurally Sound Synthetic Data

In three linesCausalSynth is a framework generating synthetic data that respects causal mechanisms in target domains. It combines a Structural Causal Model (SCM) for causal skeleton generation, an LLM as constrained realizer, and iterative consistency verification to correct structural violations. Tested on ASIA, ALARM, and MIMIC-Struct benchmarks, it achieves 96% realizability with false-positive rates near α=0.05.

Summary generated by Claude — human-verified