ECI2D Uses Causal Intervention to Distill Faithful ICL Demonstrations

The ECI2D method addresses observational equivalence in In-Context Learning by filtering demonstrations through executable causal intervention. By abstracting rationales into programs, it ensures demonstrations possess transfer invariance rather than relying on spurious post-hoc correlations.