Adam's Law：用高频词写Prompt效果更好
TL;DR · AI Summary
使用高频词汇写Prompt可显著提升模型表现，这是由FaceMind团队实验验证的Adam’s Law。
Key Takeaways
- FaceMind团队通过100种语言实验发现，高频表达方式能显著提升Prompting和Fine-tuning效果。
- Adam’s Law指出，高频词汇并非简化，而是让模型在熟悉的概率空间中工作，效果更佳。
- 写Prompt时应优先考虑模型训练语料中的词汇频率，而非追求词汇高级或优雅。
Outline
Jump quickly between sections.
Mindmap
See how the topics connect at a glance.
查看大纲文本（无障碍 / 无 JS 友好）
- Adam’s Law
Highlights
Key sentences worth saving and sharing.
FaceMind团队用100种语言、四大核心任务的实验直接证明：高频表达方式显著提升模型表现。
Adam’s Law指出，高频表达不是简化，而是让模型在最熟悉的概率空间里干活，效果直接起飞。
下次写Prompt时，别再追求多高级、多优雅了，先问自己一句：这句话模型在训练语料里见过多少次？
Everyone is still trying hard to write prompts in an elegant, professional, and tightly structured way, thinking that this will make the model more obedient and produce more accurate outputs. But the result is exactly the opposite." / X
Stop using fancy words and obscure terms with AI!! Everyone is still trying hard to write prompts in an elegant, professional, and tightly structured way, thinking that this will make the model more obedient and produce more accurate outputs. But the result is exactly the opposite. The FaceMind team directly proved through experiments in 100 languages and four core tasks: under the premise of unchanged semantics, using expressions that appear more frequently in the pre-training corpus will significantly improve the model's performance, whether in Prompting or Fine-tuning. This is Adam’s Law — the Text Frequency Law. It adds the missing fourth dimension to the current "quality-scale-difficulty" iron triangle of data engineering: frequency. High-frequency expressions are not "simplifications," but rather they allow the model to work in the probability space it is most familiar with, leading to a direct improvement in performance. Next time you write a prompt, don't aim for how advanced or elegant it is. Ask yourself first: how many times has the model seen this sentence in its training data?
Quote
Berryxia.AI
@berryxia
5h
The technical issue that "Ma Jiaqi" and Anthropic made the whole internet aware of was solved by the CEO of this startup while he was taking a shower.
Everyone is still trying hard to write prompts in an elegant, professional, and tightly structured way, thinking that this will make the model more obedient and produce more accurate outputs. But the result is exactly the opposite. ...