Beam Search · AI.WIKI

01 · 岔路口别太早封路01 · Do not close every side road too early

导航每个路口都只选眼前最快的方向，最后一定能走到总路程最短的路线吗？If navigation always chooses the fastest-looking turn, is the full trip guaranteed to be shortest?

02切换 beam 1 / 2 / 4，比较保留树与计算量Compare Retained Trees and Compute at Beam 1 / 2 / 4

高阶 · 路径联合分数expert · joint path score

状态 AMetric A

—

状态 BMetric B

—

状态 CMetric C

—

所有概率、得分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。All probabilities, scores, and outputs are fixed-seed teaching simulations, not real model output. Move the main variable back to verify the conclusion.

03把刚才的变化拆成 4 个零件Break the Change into Four Parts

Beam Search

BRANCH

候选分支Candidate branches

每一步扩展当前保留的所有序列。Expands every sequence currently retained at each step.

BEAM

保留宽度Retention width

beam 决定每层最多留下几条路径。Beam sets the maximum number of paths kept per layer.

SCORE

联合得分Joint score

比较整段 token 概率组合，不只看一步。Compares scores across the complete token sequence, not one step.

COST

计算代价Compute cost

宽度越大，同步扩展和内存开销越高。A wider beam requires more expansions and memory.

04最小机制：主交互能逐项验证Minimal Mechanism: Verify Each Link in the Interaction

MECHANISM

kept_step = top_beam(partial_sequence_score)

主交互固定同一棵玩具树；beam 只改变每层保留数量，计算量条因此同步增长。The toy tree stays fixed. Beam changes only retained paths per layer, so the compute bar rises with it.

扩展Expand 每条路径接候选add candidates to each path

排序Rank 按部分序列得分by partial-sequence score

截留Retain 只留 beam 条keep only beam paths

05把 beam 退到 1，局部赢家提前封死另一条路Reduce Beam to 1 and Prune the Better Full Route

主动制造失败Create a failure

边界实验Boundary experiment

宽度 1 每层只留一条路，等同局部贪心。紫色路径开头不是第一，却有更高整句得分。Width 1 retains a single path at every layer, equivalent to local greediness. The violet route starts behind but ends with a higher full-sequence score.

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.

✗ beam 越宽，输出一定越好。A wider beam always produces better output. → ✓ 它可能提高搜索得分，也会增加成本并偏爱通用句。It may raise search score, but costs more and can favor generic phrasing.

✗ beam 2 就会保留所有可能。Beam 2 preserves every possibility. → ✓ 每层仍会剪枝，只是比 beam 1 多留一条。Pruning still happens at every layer; it merely keeps one more path.

✗ 最终得分等于最后一个 token 的得分。Final score is the last token score. → ✓ 它来自整段序列的联合或累计得分。It comes from a joint or accumulated score across the sequence.

06 · 一句话带走06 · One line to keep

束搜索用 beam 同时保留多条部分序列，能避开局部最优陷阱，但宽度越大，扩展计算量也越高。Beam search retains multiple partial sequences to escape local optima, but wider beams require more expansion work.