CERTID Pipeline Detects Unsound Causal Identification in Reasoning Models
October 5, 2026
The CERTID pipeline uses the ID algorithm to verify if reasoning models correctly identify causal effects from observational data. It identifies failures where models provide non-identifiable formulas that cannot be validated by any observational data, using structural causal models for verification.
HOW THIS AFFECTS YOU
●
researcherYou can use this to evaluate whether reasoning models are hallucinating causal relationships that are mathematically impossible to recover.