PART 01
候选分布 Candidate distribution
每一步都有一组下一 token 概率，线越粗表示当前概率越高。 Each step has a next-token distribution; thicker lines mean higher current probability.
01 · 走迷宫时只选眼前最宽的路
01 · Always taking the widest-looking corridor
每个路口都走最宽的通道，为什么最后可能绕到一条更差的终点？ If you take the widest corridor at every junction, why can you still end at a worse destination?
02 沿最粗枝走两步，再看整体概率 Follow the Thickest Branch for Two Steps, Then Compare
局部 ≠ 整体 Local ≠ global
状态 A
Metric A
—
—
状态 B
Metric B
—
—
状态 C
Metric C
—
—
概率与输出均为固定种子的教学模拟，不代表真实模型。拖回主变量即可复核结论。 Probabilities and outputs are fixed-seed teaching simulations, not real model output. Move the main variable back to verify the conclusion.
03 把变化拆成 4 个零件 Break the Change into Four Parts
Greedy DecodingPART 02
局部 argmax Local argmax
贪心只看当前一层，立刻选择概率最大的分支。 Greedy looks only at the current layer and immediately chooses its maximum.
PART 03
追加上下文 Append to context
选中的 token 加入前缀，下一步分布由这条路径决定。 The chosen token joins the prefix and determines the next distribution.
PART 04
联合概率 Joint probability
整条路径要乘每一步概率；另一条首枝较细的句子可能整体更高。 A full path multiplies step probabilities; a thinner first branch may win overall.
04 最小机制：局部最大化不等于路径乘积最大 Minimal Mechanism: Local Maxima Do Not Maximize the Path Product
Greedy Decodinggreedy: tn = argmax p(t | prefix)
示例中 0.6×0.3 = 0.18，小于另一条 0.4×0.8 = 0.32。
In the example, 0.6×0.3 = 0.18, below the alternative 0.4×0.8 = 0.32.
每步取最大
Take each maximum
走当前最粗枝
follow the thickest branch
立即追加
Append immediately
其他分支被丢弃
other branches are discarded
计算整句
Score full path
局部赢家可能落后
local winner may lose
05 边界实验：隐藏整句联合概率 Boundary Test: Hide Full-Sequence Probability
主动制造失败 Create a failure只报每一步，误称整句最优 Report each step and falsely claim global optimality
运行贪心但遮住两条路径的概率乘积。只看到第一步 0.60 时，很容易把局部选择误报成整句最好。 Run greedy while hiding both path products. Seeing only 0.60 at step one makes it easy to misreport a local choice as the best sequence.
尚未执行。先在主交互中观察正常机制。
Not run yet. Observe the normal mechanism in the main interaction first.
✗
每步最高，所以整句概率也最高。
Highest at each step means highest full-sequence probability.
→ ✓
被丢弃分支后续可能拥有更高条件概率。
A discarded branch may have much stronger later probabilities.
✗
贪心等于低温采样。
Greedy is the same as low-temperature sampling.
→ ✓
贪心不抽样；它直接取 argmax。
Greedy does not sample; it directly takes argmax.
✗
确定性输出就一定更准确。
Deterministic output must be more accurate.
→ ✓
稳定只说明可重复，不保证内容质量。
Stability means repeatability, not guaranteed quality.
06 · 一句话带走
06 · One line to keep