Predictive gains, causal losses: why interventional competence cannot be learned from data alone

| Source: arXiv AI

Tags: Causal Inference, Interventions, Theory, Machine Learning, Generalization

A theory paper proves that any predictive gain beyond a model's direct causal inputs forces a positive lower bound on its error under interventions, arguing interventional competence cannot come from data alone.

Details

The paper studies deterministic predictors under squared-error loss. If a model predicts better than the best predictor using only an outcome's direct causes, it must depend on other inputs, and when interventions change those inputs without changing the outcome, that dependence yields a nonzero floor on intervention-response error. It also identifies conditions where gradient updates introduce this dependence or damage a previously correct mechanism through shared parameters while improving other predictions. The author argues that error below this floor requires admitting better causal hypotheses, not more data, and conjectures that keeping interventional competence needs autonomous participation in proposing and testing alternatives.