Maximum Output Length

01 · 广播被定时器切断 01 · A broadcast cut by a timer

一句话停在逗号后，是作者自然写完，还是计时器突然拔了插头？ If a sentence stops after a comma, did the writer finish—or did a timer pull the plug?

02 拖长度上限，区分硬截断与自然 EOS Move the Limit and Distinguish Hard Cut from Natural EOS

两种停止标记 Two stop labels

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

Maximum Output Length

PART 01

输出计数 Output counter

只计算本次新生成的 token，不等同于输入上下文长度。 It counts newly generated tokens, not the input context length.

PART 02

长度上限 Length limit

应用在生成循环外部，到格数就强制停止。 It is applied outside the generation loop and forces a stop at the cap.

PART 03

EOS EOS

模型在候选中选择结束 token，属于自然停止机制。 The model selects an end token from candidates, producing a natural stop.

PART 04

停止标记 Stop label

系统必须区分“被截断”与“自然结束”，下游才能正确处理。 Systems must distinguish “truncated” from “natural EOS” so downstream logic can act correctly.

04 最小机制：谁先到，就由谁触发停止 Minimal Mechanism: Whichever Arrives First Stops Generation

Maximum Output Length

stop = EOS selected ∨ generated_count ≥ max_tokens

本页把自然 EOS 固定在第 6 格；长度上限若小于 6，就明确标记被截断。 This page fixes natural EOS at slot 6. Any limit below 6 is explicitly labeled truncated.

生成一格 Generate a slot 计数加 1 counter increments

比较两边界 Compare boundaries EOS 或上限先到 EOS or limit arrives first

写停止原因 Record reason 自然结束或被截断 natural or truncated

05 边界实验：在 EOS 前把生成带剪断 Boundary Test: Cut the Strip Before EOS

主动制造失败 Create a failure

短，不代表完整 Short does not mean complete

把最大输出长度设为 3，而自然 EOS 在第 6 格。生成带被外部剪刀切断，必须显示“被截断”。 Set maximum output length to 3 while natural EOS sits at slot 6. The external scissors cut the strip and must label it truncated.

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.

✗ 达到最大长度就是自然说完。 Reaching max length means a natural finish. → ✓ 这是外部强制停止，可能截在半句。 It is an external forced stop and may cut mid-sentence.

✗ 最大输出长度等于上下文窗口。 Maximum output length equals context window. → ✓ 一个限制新输出，一个限制一次计算可接收范围。 One caps new output; the other caps what one computation can receive.

✗ 只要提高上限，回答就一定更好。 A higher cap always improves answers. → ✓ 更长只提供空间，也会增加成本和跑偏机会。 More room also increases cost and room to drift.

06 · 一句话带走 06 · One line to keep

最大输出长度是外部强制边界；拖到 EOS 之前会显示“被截断”，让它与自然选择结束 token 明确分开。 Maximum output length is an external hard boundary. Move it before EOS to see “truncated,” clearly separated from selecting the end token naturally.