「When engaging in multi-page reading comprehension, humans typically begin by identifying the pages likely to contain the answer, and then focus on locating the specific regions that correspond to the question and answer within those pages. Inspired by this “coarse-to-fine” reading strategy, EviGRPO mimics the human approach by first selecting a small set of potentially relevant pages at a coarse level, followed by fine-grained reasoning over the selected content.」とのことだが、このようなドメイン（タスク）特化のアプローチはいまだ有効なんだろうか。。