中文直达

English first, Chinese below. The English is translated from the author's Chinese. 英文在前，中文在后。英文由作者的中文原稿翻译而来。

Shortly after we released the WORTH Benchmark paper, an independent researcher in Europe wrote to us. He asked about how the key clue sentences were designed. It is a good question, and the paper answers it in only one sentence. So here is a longer answer.

The clues come from the case, not from the text

We do not write a story first and then pick clue sentences from it. The order is the other way around.

Our creative team has three people: me, my co-author Chao Xu, who is a novelist, and one colleague. Chao Xu first designs the core trick of the case: how the crime is done, and how it is hidden.

Then the three of us discuss it again and again. If we do this, what happens? What if we do that instead?

We stop only when we are fully sure that two things make complete sense: how the suspect commits the crime, and how the solver finds the clues and reaches the truth.

When this logic is fixed, the clues a reader needs are basically fixed too. Chao Xu then writes the story and turns these clues into specific sentences. So the key clues exist before the story text does.

Earlier in this series: The Dog That Did Not Bark: Can an LLM Find the Clues Nobody Points Out?

WORTH 札记 #1：哪些句子是线索，由谁决定？

Back to English / 返回英文

WORTH Benchmark 论文发布后不久，一位欧洲的独立研究者给我们来信，询问关键线索句的设计信息。这是个好问题，论文里只用一句话带过，所以在这里展开说说。

线索来自案件，不是来自文本

我们不是先写好故事，再从里面挑线索句。顺序正好相反。

我们的创作团队有三个人：我、论文的另一位作者徐超（小说家），还有一位同事。先由徐超设计案件的核心诡计：凶手怎么作案，又怎么把它藏起来。

然后三个人开会反复推演：如果这样，会怎样？换成那样呢？

等到我们完全确定了，「嫌疑人作案经过」以及「破案者获得线索并破获真相」，这两件事都完全合理，我们就停止讨论。

这套逻辑定下来，读者破案需要的线索也就基本定了。之后由徐超执笔写故事，把这些线索落成具体的句子。所以，关键线索在故事文本存在之前就已经存在。

前一篇：一条没叫的狗：大模型能找到没人指给它的线索吗？

Paper on Hugging Face / 论文页：WORTH Benchmark