A new arXiv paper introduces partial-revelation panel forecasting, a problem where a model observes only k of n annotator votes on a previous dialogue turn and must predict the complete count distribution of the next turn’s panel. This moves beyond single-label prediction to a distribution-level forecast that accounts for the unrevealed portion of the panel.
The authors frame the task around cross-turn predictive information, examining how partial observations on one turn inform the full panel of the following turn. They also propose coarsening-consistent operators, which are designed to behave consistently as the revealed subset changes in size or composition.
Because the abstract is truncated, full details of the proposed operators and experimental results are not available from the source. The key contribution as described is a formal treatment of partial-revelation forecasting for panel distributions.