Controlled Acquisition and Abstention in Three-Channel Score Conflicts
The paper tests when a model should pay for a missing modality or abstain when audio, video, and text scores conflict.
It frames the problem as a controlled three-score benchmark where a policy sees two scores, can request the third at a cost, or abstain. A threshold policy matched always-request decisions while using fewer requests, but its value depended on how ambiguity was rewarded. In synthetic tests, the threshold policy beat a three-score majority reference on targeted decision accuracy and utility, though the majority used more information. Matched-budget routing gains were mixed, and scoring all abstentions as wrong made majority stronger in utility. Source: HF Daily Papers' note.
It frames the problem as a controlled three-score benchmark where a policy sees two scores, can request the third at a cost, or abstain. A threshold policy matched always-request decisions while using fewer requests, but its value depended on how ambiguity was rewarded. In synthetic tests, the threshold policy beat a three-score majority reference on targeted decision accuracy and utility, though the majority used more information. Matched-budget routing gains were mixed, and scoring all abstentions as wrong made majority stronger in utility. Source: HF Daily Papers' note.
score 3