Learning Deployable Causal Action Geometry under Temporal Non-Stationarity
The paper addresses temporal non-stationarity where outcome levels drift over time despite stable causal effects. It proposes a 'deployable causal action geometry'—a stable action-response shape—as the decision target instead of raw predictions. This allows systems to make optimal decisions by focusing on the persistent structure of how actions affect responses, rather than fluctuating outcome magnitudes.