Interventional Data Fails to Correct Causal Direction in LLMs
August 3, 2026
Models trained with interventional data still default to observational correlations in Simpson's-paradox scenarios. In 58% of tested worlds, purely observational contexts induce systematic sign reversals regardless of the training mixture, suggesting inference-time context outweighs pretraining data type.
HOW THIS AFFECTS YOU
●
researcherYou should focus on inference-time context management rather than just increasing interventional data during pretraining to fix causal reasoning.