Lost in the Middle

01 · 同一张便签，夹在中间就消失了01 · Same note, invisible in the middle

证据一个字都没改，只从第一段移到第五段；如果回答突然漏掉它，问题来自内容，还是位置？The evidence text is unchanged and moves only from paragraph one to paragraph five. If the answer misses it, is content—or position—the cause?

02拖动同一证据到开头、中间与结尾，观察命中率热图和引用位置Drag Identical Evidence to the Start, Middle, and End; Watch the Heatmap and Citation Slot

同内容 · 只改位置same content · position only

状态 AMetric A

—

状态 BMetric B

—

状态 CMetric C

—

所有概率、命中率、得分与输出均为固定种子的教学模拟，不代表真实模型。拖回主变量即可复核结论。All probabilities, hit rates, scores, and outputs are fixed-seed teaching simulations, not real model output. Move the main variable back to verify the conclusion.

03把刚才的变化拆成 4 个零件Break the Change into Four Parts

Lost in the Middle

EVIDENCE

同一证据Same evidence

文本、长度与事实保持完全一致。Text, length, and facts remain identical.

POSITION

上下文位置Context position

只在 1–9 段之间移动。It moves only among slots 1–9.

HEATMAP

命中率热图Hit-rate heatmap

首尾较亮、中部较暗的教学分布。A teaching distribution brighter at the edges and darker in the middle.

CITATION

引用槽位Citation slot

回答引用位置随证据同步移动。The answer's citation slot moves with the evidence.

04最小机制：主交互能逐项验证Minimal Mechanism: Verify Each Link in the Interaction

MECHANISM

hit_rate(position) ≈ high at edges, low near middle

所有变量都固定，只有 position 改变；因此热图差异专门展示位置效应，而不是内容质量差异。Every variable is fixed except position, so the heatmap isolates ordering effects rather than content quality.

固定证据Fix evidence内容不变content unchanged

改变位置Move position开头 / 中间 / 结尾start / middle / end

测量利用Measure use命中率 + 引用槽hit rate + citation slot

05把证据拖到正中间第 5 段Drag the Evidence to the Center, Slot 5

主动制造失败Create a failure

边界实验Boundary experiment

如果只检查证据是否存在，会错过位置造成的利用下降。点击后只改位置，不改任何文字。Checking only whether evidence is present misses position-related degradation. This changes location without changing any text.

尚未执行。先在主交互中观察正常机制。Not run yet. Observe the normal mechanism in the main interaction first.

✗ 证据在窗口里就会被同等利用。Evidence inside the window is used equally.→ ✓ 位置可能影响利用率，尤其在长上下文中。Position can affect utilization, especially in long contexts.

✗ 把所有关键证据都放开头即可。Put every key fact at the start.→ ✓ 关键材料仍需排序、去重与真实评测。Critical material still needs ordering, deduplication, and real evaluation.

✗ U 形热图是每个模型的固定规律。Every model has the same U-shaped curve.→ ✓ 本页是教学模拟，实际强度依模型与任务而变。This is a teaching simulation; strength varies by model and task.

06 · 一句话带走06 · One line to keep

中间信息遗失说明证据存在不等于被同等利用；保持内容不变并移动位置，才能把排序效应测出来。Lost in the middle shows that presence does not guarantee equal use. Keep content fixed and move its position to measure ordering effects.