Extend acquisition guarantees to adaptive iterative feedback selection

Extend the theoretical guarantees for Welfare-Aware Preference Acquisition (WAPA) to iterative acquisition procedures that adapt to observed human preference labels and updated policy parameters.

Background

WAPA's exact optimality result applies to a fixed estimated acquisition-welfare objective with frozen model predictions, while the paper's acquisition protocol selects queries before their labels are revealed. The paper also gives a conditional generalization bound under assumptions including label-independent acquisition.

The unresolved problem is to establish corresponding guarantees when acquisition proceeds iteratively: newly observed labels affect subsequent query choices, and the downstream policy is updated during the acquisition process. Such an extension would address the gap between the paper's static, frozen-prediction setting and adaptive human-feedback collection.