关于「接力保留率连乘:层数上限、滚动摘要的近因斜坡与细节的伤亡」,《中段遗忘与接力阈值》给出的定量结论是什么?

结论:第四章计算接力保留率连乘,说明滚动摘要把 U 形变成近因斜坡,细节类事实八层后只剩 0.4305。

数据口径:书中任务、材料与参数均为示意,模型常数来自 2023 年的一个模型;全部计算由书中程序生成,不调用任何模型接口;本书是非个人化经营建议,不评测、不推荐任何模型或服务。 书中以标注框区分已发表的证据、可运行的 Python 计算(21 段程序)、原典与作者的假说。

English version

对照数据表

下表数字由本章 Python 程序原样打印,可重新运行复现。 程序只用 Python 标准库与结论吻合的打印值: 0.4305

mean per-layer retention E r = 0.9475
L(E r)^LE[r^L]Monte Carlodetails onlysalient only
10.94750.94750.94690.90000.9950
20.89780.90000.89900.81000.9900
40.80600.81810.81950.65610.9801
80.64960.69560.69530.43050.9607
160.42200.55410.55360.18530.9229
after 8 layers details are 0.309 of the surviving facts

核心公式

\[R_L = r^L .\]
查看完整验算程序
# Chapter 4. p43: when facts differ, the average hides the casualties.
# Illustrative mixture: half the facts are salient (kept with chance
# 0.995 per layer), half are details such as figures, names and dates
# (kept with chance 0.90). A summary is scored by its average r, but
# survival after L layers is E[r^L], which by Jensen's inequality is
# at least (E r)^L, while the detail class alone decays as 0.90^L.
# Fixed-seed Monte Carlo of 20000 facts checks the exact numbers.
import math
import random

CLASSES = ((0.5, 0.995), (0.5, 0.90))
FACTS = 20000
rng = random.Random(20261113)

mean_r = math.fsum(w * r for w, r in CLASSES)
print("mean per-layer retention E r = %.4f" % mean_r)
print("  L  (E r)^L   E[r^L]  Monte Carlo  details only  salient only")
for depth in (1, 2, 4, 8, 16):
    exact = math.fsum(w * r ** depth for w, r in CLASSES)
    alive = 0
    for j in range(FACTS):
        r = CLASSES[0][1] if j % 2 == 0 else CLASSES[1][1]
        if all(rng.random() < r for _ in range(depth)):
            alive += 1
    print("%3d  %7.4f  %7.4f  %11.4f  %12.4f  %12.4f"
          % (depth, mean_r ** depth, exact, alive / FACTS,
             CLASSES[1][1] ** depth, CLASSES[0][1] ** depth))
share = 0.5 * 0.90 ** 8 / math.fsum(w * r ** 8 for w, r in CLASSES)
print("after 8 layers details are %.3f of the surviving facts" % share)

唯一数据源与引用

范雅福(Fan, Y.). (2026). 《长上下文注意力中段遗忘曲线、位置偏置与零损耗分段摘要接力阈值母典》第 4 章「接力保留率连乘:层数上限、滚动摘要的近因斜坡与细节的伤亡」(版本 1.0.0). 5D Library. DOI 待发布

DOI 待发布 · 专著页正式发布后上线(本站暂无专著页) · 5DL-B-255 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026255,
  author    = {Fan, Yafu},
  title     = {Mid-Context Forgetting in Long-Context Attention: Positional Bias and the Threshold for Lossless Segment-Summary Relays},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

本书其他章节结论