Downstream mechanisms of target-support gains

Identify which downstream components of ALIGNNN—its message-passing layers, readout layer, or their joint adaptation to target-containing environments—account for the portion of target-support improvement that remains after output calibration and freezing the initial element-projection layer.

Background

The paper shows that adding a small number of labeled target-containing structures substantially improves formation-energy predictions for held-out elements. Calibration experiments indicate that constant or affine corrections to zero-shot outputs explain only part of this improvement, while frozen-stem retraining preserves most of the gain for five of six evaluated elements. These results rule out output correction and relearning of the initial element projection as complete explanations.

The remaining unexplained improvement could result from adaptation in the message-passing layers, the readout layer, or coordinated changes across these components. The authors explicitly state that their present controls do not distinguish among these possibilities, leaving the responsible downstream mechanisms unresolved.