PART 01
当前上下文 Current context
开头提示与所有已生成 token 拼在一起。 The prompt and every generated token are joined together.
01 · 先想一条接龙
01 · Think of a word chain
如果第二个人把“整理花园”改成“穿过大门”，后面的人还能沿用原来的故事吗？ If the second person changes “tending the garden” to “crossing the gate,” can everyone after them keep the old story?
02 因果分叉台：固定随机带，只换一个 token Causal Fork Bench: Hold Randomness Fixed, Replace One Token
关键页 · 误差传播 key lesson · error propagation
已生成
Generated
0/4
逐 token
token by token
分叉点
Fork point
—
只强制改一项
one forced change
后续受影响
Downstream changed
0
不同 token 数
different tokens
03 一个自回归回路的 4 个零件 Four Parts of the Autoregressive Loop
xt → contextt+1PART 02
下一轮分布 Next distribution
当前上下文决定这一轮候选及概率。 Current context determines this round's candidates and probabilities.
PART 03
选中 token Selected token
固定随机带给出本轮抽样位置。 The fixed random tape supplies this round's draw position.
PART 04
Feedback
选中项被送回输入，成为下一轮的新条件。 The selected item feeds back into the input as the next condition.
04 最小机制：同一随机带，为什么后续仍会变 Minimal Mechanism: Why the Future Changes on the Same Random Tape
P(xt+1 | x≤t)xt ≠ x′t ⇒ Pt+1 ≠ P′t+1
本页反事实分支沿用完全相同的随机数序列，只在所选步骤强制换一个 token；因此后续差异来自上下文变化。
The counterfactual branch reuses the exact same random sequence and forces only the chosen token, so downstream differences come from changed context.
换掉一步
Replace one step
新 token 入栈
new token enters context
下一轮分布改变
Next distribution changes
同一随机数落到别处
same draw lands elsewhere
整枝分叉
The branch diverges
差异继续回灌
difference keeps feeding back
05 边界实验：第一步就走另一条路 Boundary Test: Take the Other Path at Step One
主动制造失败 Create a failure越早的偏差，暴露给越多后续步骤 Earlier deviations reach more downstream steps
强制替换第一步，同时保持随机带不变。比较基准线与红色分支有多少 token 不同。 Force an alternate first token while holding the random tape fixed. Compare how many tokens differ between the baseline and red branch.
尚未执行。分叉越早，可被改写的后续越长。
Not run yet. The earlier the fork, the longer the continuation exposed to change.
✗ 每一步预测彼此独立。
Each prediction step is independent.
→ ✓ 上一 token 会进入下一轮条件。
The previous token enters the next round's condition.
✗ 修正一个词只会改那一个词。
Correcting one token changes only that token.
→ ✓ 修正后必须重新生成所有后续。
Every downstream token must be regenerated afterward.
✗ 固定种子保证手动改词后仍有相同续写。
A fixed seed guarantees the same continuation after a manual edit.
→ ✓ 种子固定随机数，不固定已改变的条件分布。
The seed fixes random numbers, not a distribution changed by context.
06 · 一句话带走
06 · One line to keep