Research Style

我的科研思路主要回答两个问题：好问题从哪里来？问题确定以后，方法从哪里来？

My research style centers on two questions: Where does a good problem come from, and how should the method follow once the problem is defined?

01Back to the Task回到任务本身Return to the source
→
02Long-term Vision提出清晰的理想愿景Define the ideal destination
→
03Define the Problem定义具体问题Make the question precise
→
04Derive the Method由问题推导方法Let the problem constrain the solution
01

紧盯一手资料

Back to the Task

Back to the Task

信息传播会带来信息衰减。研究不能只停留在别人对任务的解释上，而要穿过论文、综述和总结，重新回到任务本身。

Information decays as it propagates. Research should not stop at other people's interpretations of a task; it should trace through papers, surveys, and summaries and return to the task itself.

信息传播与衰减Information propagation and decayCloser to the source → higher task fidelity
100%任务本身Task itselfSource
→
82%开坑之作Pioneering work1st interpretation
→
63%后续论文Follow-up worksRefinement
→
43%SurveyAbstraction
→
24%Blog / PPTSummary
资料越多，越要回到源头。
The more information we have, the more important it is to return to the source.
VISION

回到任务的一个简单高效的方式，是为任务提出一个清晰、明了的理想愿景。

A simple and effective way to return to the task is to formulate a clear, explicit vision of its ideal form.

当我们知道理想情况应该是什么样，就可以从终点向今天倒推：哪些还没有做好？差距在哪里？现在最应该解决什么？

Once the ideal destination is clear, we can reason backward to the present: What is still missing? Where is the gap? What should we solve now?

任务本身Task→Ideal Vision→Work Backward→Gap→What to Do Now
Example · AIGC Forensics从今天的任务定义，看向真正的远期目标From today's formulation to the long-term vision
TODAY
▧
Image↓Detector↓Real / Fake

把二分类做得更准

Improve binary detection

GAP→从理想愿景倒推Work backward from the vision
LONG-TERM VISION
▧
Image↓
Source来源Anomaly异常Location位置
↓What? · Where? · Why?

判断 + 定位 + 解释 + 证据

Detection + localization + explanation + evidence

先把理想终点说清楚，再向今天倒推。这样，研究问题来自理想愿景与当前能力之间的差距，而不是来自“还能换什么方法”。

Make the ideal destination explicit, then reason backward. Research problems emerge from the gap between the vision and current capability, rather than from asking what method to try next.

02

现象、问题与方法

Phenomenon, Problem, and Method

Define the Problem, Derive the Method

Problem>Method

现象不是问题。先拆出元素，再研究元素的特性，才能把现象收敛成一个具体问题。问题一旦足够清楚，就会对解空间施加强约束，方法因此可以从问题中被“推导”出来。

A phenomenon is not a problem. Decompose the phenomenon into elements and properties, then narrow it into a precise question. A well-defined problem strongly constrains the solution space, allowing the method to be derived rather than arbitrarily chosen.

01现象PhenomenonWhat happened?
→
02元素ElementsWhat is involved?
→
03特性PropertiesWhat matters?
→
04问题ProblemWhat exactly to answer?
→
05方法MethodHow to answer it?
直观例子 · “树叶落了”只是现象Intuitive example · “Leaves are falling” is only a phenomenonDefine the Problem → Derive the Method
🍂
树叶落了Leaves are fallingPHENOMENON ≠ PROBLEM
↓
🌳 树Tree🍃 叶Leaf🌦 环境Environment◷ 时间Time
↓ PROPERTIES
💧 水分Water🌡 温度Temperature☀ 光照Light
↓ DEFINE THE PROBLEM
PROBLEM A是否由缺水导致？Is water deficiency the cause?↓改变浇水量Vary water supplyMETHOD A
PROBLEM B是否由低温导致？Is low temperature the cause?↓改变温度Vary temperatureMETHOD B
PROBLEM C是否由光照不足导致？Is insufficient light the cause?↓改变光照Vary illuminationMETHOD C

同一个现象可以对应多个问题；不同的问题自然导出不同的方法。不是“树叶落了，所以设计一个方法”，而是先明确“为什么落”。

The same phenomenon can lead to different problems, and different problems naturally imply different methods. Do not start by inventing a method for falling leaves; first define what cause you are testing.

METHOD FIRST

看到现象 → Attention？Agent？新 Loss？

Phenomenon → Attention? Agent? New loss?

↓

哪个有效用哪个

Keep whatever works

DEFINE↓问题约束解空间Problem constrains solution space↓DERIVE
PROBLEM FIRST

Phenomenon → Elements → Properties

↓

Problem → Method

问题定义得越清楚，方法越不是“想出来”的，而是“推出来”的。The clearer the problem, the less the method is “invented” and the more it is “derived.”A well-defined problem constrains its solution.

我的研究路径

My Research Path

Back to the Task→Long-term Vision→Work Backward→Gap→Define the Problem→Derive the Method→Evidence

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