锯齿状能力从何而来？受控两层任务研究
原标题：Where Does Jagged Competence Come From?
AI 摘要
该研究构建了一个由已知两层组成的受控任务：下层A从记录中计算五个逐槽和，上层B利用槽1的和及历史模式预测下一类别，因此B无法仅由当前A计算得出。训练小型循环网络仅学习B后发现，网络对A的习得呈现「锯齿状」——B需要的槽1可读性达97-98%，而其他槽仅5-10%，类别误读集中在槽1临界值附近。匹配实验表明，即使提供精确的A也不足以消除B的失败，说明下层理论的不均匀获取、访问与保持部分解释了上层能力的锯齿状表现。
正文节选
Where Does Jagged Competence Come From? Abstract Capable systems often show jagged competence: low average error alongside failures on particular inputs. We ask where it comes from in a task built from two known layers. A lower layer A computes five per-slot sums from records; an upper layer B uses the slot-1 sum and a mode carried over from earlier boards to predict the next board’s category, so B cannot be computed from the current A alone: B is a strict extension of A. We train small recurren