PART 01
Logits
候选 token 的未归一化分数。上方五个值固定，不随温度改变。 Unnormalized candidate scores. The five values above stay fixed as temperature moves.
01 · 先从选择困难开始01 · Start with a choice
同一组候选词，怎样让回答更稳定，又怎样让它偶尔冒出新花样？ With the same candidate words, how do you make answers steadier—or let fresh options surface?
02 概率风洞：只动温度Probability Wind Tunnel: Change Temperature Only
滑块分布型Distribution slider
最高概率Top probability
0%
—
分布熵Entropy
0.00
越高越分散higher = flatter
Logit range
3.4
始终不变always fixed
03 把滑块变化拆成 4 个零件Break the Slider Effect into Four Parts
同一组候选Same candidatesPART 02
温度缩放Temperature scaling
每个 logit 都除以同一个 T。T 小会放大差距，T 大会压缩差距。 Every logit is divided by the same T. Low T widens gaps; high T compresses them.
PART 03
Softmax
把缩放后的分数转成总和为 1 的概率分布。 Converts the scaled scores into probabilities that sum to one.
PART 04
采样Sampling
按新概率抽取下一 token。温度不是“创造力开关”，只是改变抽样权重。 Draws the next token from the new probabilities. Temperature changes weights; it is not a creativity switch.
04 最小公式：差距怎样被放大或压平Minimal Formula: Widen or Compress the Gaps
Σpi = 1pi(T) = ezi/T / Σjezj/T
z 是不变的 logit；T 是上方唯一变量；p 是画布中的彩色柱。
z is the fixed logit, T is the only variable above, and p is each colored bar.
降低 TLower T
z/T 差距放大z/T gaps widen
概率变尖Distribution sharpens
头部候选更集中Top candidates dominate
抽样更稳定Samples stabilize
重复词更常出现The same token appears more often
05 边界实验：把分布调得太平Boundary Test: Flatten the Distribution Too Far
主动制造失败Create a failure让低分候选过度抬头Overpromote weak candidates
点击把 T 拉到 2.0。观察最高概率下降、熵上升，原本很弱的候选也更常进入样本。 Set T to 2.0. Watch the top probability fall, entropy rise, and weak candidates enter the sample more often.
尚未执行。高温不保证“更有创意”，只会让分布更平。
Not run yet. High temperature does not guarantee creativity; it only flattens the distribution.
✗ 温度会修改模型参数或原始 logits。Temperature changes model weights or raw logits.
→ ✓ 它只在采样前缩放固定 logits。It only scales fixed logits before sampling.
✗ 温度越高，答案质量一定越高。Higher temperature always improves quality.
→ ✓ 高温增加多样性，也会提高弱候选被抽中的机会。High T adds variety and also raises the odds of weak candidates.
✗ 低温等于完全确定。Low temperature is fully deterministic.
→ ✓ 只要还在采样，非头部候选通常仍有非零概率。While sampling remains enabled, non-top candidates usually retain nonzero probability.
06 · 一句话带走06 · One line to keep