Why is this so hard? Models must identify relevant sections of input, label or categorize these sections, and then accumulate information to make distributional-level decisions. Adding labels in-context or specifying more reasoning effort has limited benefit.
