关于「反方与异质:四种曲线、更大的窗口与单针测试」,《中段遗忘与接力阈值》给出的定量结论是什么?
结论:程序 p21 显示原文四个模型画出四种形状,只有一个模型的中段低于闭卷成绩,而所有模型都有明显的稀释损失。程序 p22 显示扩展窗口版本与原版的差距不超过 0.3 个百分点,远小于抽样误差。程序 p23 说明单针召回 0.997 换到一百条事实上只剩 0.740 的全中机会。描述到此为止已经够了;要预测一个从未测过的长度,我们需要一个能外推的公式。首因、近因与稀释,能不能写进同一条曲线?
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。
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: 0.3, 0.740, 0.997
| per-fact p all of 1 all of 5 all of 20 all of 100 k for 0.5 | |||||
| 0.997 | 0.997 | 0.985 | 0.942 | 0.740 | 230.7 |
| 0.990 | 0.990 | 0.951 | 0.818 | 0.366 | 69.0 |
| 0.950 | 0.950 | 0.774 | 0.358 | 0.006 | 13.5 |
| 0.900 | 0.900 | 0.590 | 0.122 | 0.000 | 6.6 |
| expected misses among 100 facts at p = 0.997: 0.3 | |||||
查看完整验算程序
# Chapter 2. p23: why a near-perfect single-needle score settles little.
# A single-needle test reports the chance p that one planted fact is
# found. A real task needs k facts at once. If misses were independent,
# all k survive with p^k; we also print the k at which that falls to
# one half. The value 0.997 echoes the single-needle recall above
# 99.7 per cent reported for Gemini 1.5 Pro (Gemini Team 2024); the
# other values are illustrative. Independence is an assumption: shared
# failure modes would raise p^k for some tasks and lower it for others.
import math
PS = (0.997, 0.99, 0.95, 0.9)
KS = (1, 5, 20, 100)
print("per-fact p " + " ".join("all of %3d" % k for k in KS)
+ " k for 0.5")
for p in PS:
row = " ".join("%10.3f" % (p ** k) for k in KS)
k_half = math.log(0.5) / math.log(p)
print(" %6.3f %s %8.1f" % (p, row, k_half))
print("expected misses among 100 facts at p = 0.997: %.1f"
% (100 * (1 - 0.997)))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 2 章「反方与异质:四种曲线、更大的窗口与单针测试」(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}
}