RESEARCH PLAYBOOK

我如何做研究

How I Do Research

两个原则：回到任务本身，找到真正值得做的问题；定义问题，再推导方法。

Two principles: go back to the task to find the right problem, then define the problem and derive the method.

Back to the Task→Ideal Vision→Gap→Problem→Method

BACK TO THE TASK

不要沿着论文往后追，要沿着任务往回看。

Don't follow the papers. Follow the task.

论文会不断告诉我们“别人现在怎么做”。但一个简单高效的方式，是暂时放下现有方法，先为任务提出一个清晰、明了的理想愿景。

Papers keep telling us how others currently solve a task. A simple alternative is to temporarily set methods aside and formulate a clear ideal vision of what the task should ultimately become.

沿着已有论文Follow the papers

Paper A→Paper B→Paper C→+0.8%

局部优化Local optimization

回到任务本身Return to the task

THE TASK→Ideal Vision→What is missing?→NEW PROBLEM

从理想终点倒推今天Work backward from the ideal

知道理想情况是什么样，就可以倒推：哪些还没做好？差距在哪里？现在该做什么？

Once the ideal is clear, work backward: What is missing? Where is the gap? What should we do now?

EXAMPLE · AIGC FORENSICS

TODAYImage↓Detector↓Real / Fake

真正解决这个任务，最终只需要一个二分类标签吗？

If the task were truly solved, would a binary label be enough?

IDEALImage↓

SourceWhere?What?Why?

↓Evidence

理想愿景与当前能力之间的差距，就是问题的来源。

The gap between the ideal vision and today's capability is where new problems come from.

DEFINE THE PROBLEM, DERIVE THE METHOD

现象不是问题。问题清楚，方法才清楚。

A phenomenon is not a problem. A clear problem leads to a clear method.

Phenomenon ≠ ProblemProblem → MethodProblem > Method

🍂树叶落了Leaves are fallingPHENOMENON

↓

💧 水分Water🌡 温度Temperature☀ 光照Light

DEFINE THE PROBLEM ↓

PROBLEM A是否缺水？Water deficiency?↓改变浇水量Vary waterMETHOD A

PROBLEM B是否低温？Low temperature?↓改变温度Vary temperatureMETHOD B

PROBLEM C是否缺光？Insufficient light?↓改变光照Vary lightMETHOD C

METHOD FIRST

看到现象 → Attention？Agent？新 Loss？

Phenomenon → Attention? Agent? New loss?

PROBLEM FIRST

Phenomenon → Elements → Properties → Problem → Method

问题越模糊，方法越多；问题越清楚，方法越确定。The vaguer the problem, the larger the method space. The clearer the problem, the more constrained the method.

THE WHOLE PLAYBOOK

Task→Ideal Vision→Gap→Phenomenon→Elements & Properties→Problem→Method→Evidence

Back to the Task. Define the Problem. Derive the Method.

从任务的理想形态倒推问题，从清晰定义的问题推导方法。

Work backward from the task's ideal form to find the problem; derive the method from a precisely defined problem.