Assess viability of proposed moral-graph training approaches

Evaluate the viability of the proposed training approaches that use moral contexts to retrieve applicable values and either (i) fine-tune with adherence ratings to values cards or (ii) train a reward model on wisdom-upgrade orderings, including empirical performance and alignment outcomes.

Background

The authors outline two concrete training approaches: SFT/RLHF-like datasets rated for adherence to values cards per moral context, and reward modeling based on wisdom-upgrade orderings derived from the moral graph.

They do not yet have evidence of the viability of these approaches and explicitly flag the need for further evaluation.