关于「最优段长与接力阈值:读的损失与写的损失之间的拉锯」,《中段遗忘与接力阈值》给出的定量结论是什么?

结论:第五章求最优段长与接力阈值:对 2023 年的读者,每次摘要调用的保真度须高于约 0.911,段长须很短;读者损失低于原曲线的 0.413 倍时,保真度 0.97 的接力就输给整窗。

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: 0.413, 0.911, 0.97

lamstuffedrelay q=.99q=.97q=.95winner at q=.97
1.000.4760.9230.7840.663relay
0.500.7380.9230.7840.674relay
0.250.8690.9230.8060.764stuffed
0.100.9480.9260.8820.846stuffed
0.050.9740.9490.9110.874stuffed
0.001.0000.9800.9410.902stuffed
q = 0.99: relay wins only when lam exceeds 0.148
q = 0.97: relay wins only when lam exceeds 0.413
q = 0.95: relay wins only when lam exceeds 0.643

核心公式

\[R(s) = \big(q\, c(s)\big)^{L},\]
查看完整验算程序
# Chapter 5. p53: newer, flatter readers erase the relay's advantage.
# Later work flattens the curve (calibration, position-aware training,
# near-perfect single-needle scores). We shrink every loss of the p32
# reader by a factor lam: rho_lam = 1 - lam (1 - rho). lam = 1 is the
# 2023 GPT-3.5-Turbo shape, lam = 0 a perfectly flat reader. For
# D = 256 units and per-call fidelity q we compare the best tree relay
# (segment s from 2 to 64) with one stuffed window. Illustrative.
import math

TAU, M0, K, A, B = 2.4, 0.752, 0.0505, 0.251, 0.128
D = 256


def rho(i, n, lam):
    v = (M0 - K * math.log(n) + A * math.exp(-(i - 1) / TAU)
         + B * math.exp(-(n - i) / TAU))
    return 1.0 - lam * (1.0 - min(1.0, v))


def capture(n, lam):
    return math.fsum(rho(i, n, lam) for i in range(1, n + 1)) / n


def layers(n, f):
    depth, width = 0, n
    while width > 1:
        width = -(-width // f)
        depth += 1
    return max(depth, 1)


def best_relay(q, lam):
    cap = {s: capture(s, lam) for s in range(2, 65)}
    return max((q * cap[s]) ** layers(D, s) for s in cap)


print("  lam   stuffed   relay q=.99  q=.97  q=.95   winner at q=.97")
for lam in (1.0, 0.5, 0.25, 0.1, 0.05, 0.0):
    st = capture(D, lam)
    rs = [best_relay(q, lam) for q in (0.99, 0.97, 0.95)]
    print("%5.2f   %7.3f   %10.3f %6.3f %6.3f   %s"
          % (lam, st, rs[0], rs[1], rs[2],
             "relay" if rs[1] > st else "stuffed"))
for q in (0.99, 0.97, 0.95):
    lo, hi = 0.0, 1.0
    for _ in range(40):
        mid = (lo + hi) / 2
        if best_relay(q, mid) > capture(D, mid):
            hi = mid
        else:
            lo = mid
    print("q = %.2f: relay wins only when lam exceeds %.3f" % (q, hi))

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 5 章「最优段长与接力阈值:读的损失与写的损失之间的拉锯」(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