Both halves share one discipline:First determine whether the problem is at the level of ideas or the level of sentences, then act. Polishing sentences on a wrong skeleton is the most expensive kind of rework.
I. Revising Your Own Writing
1. The level of ideas (do this first)
For the abstract, introduction, method, or any paragraph, four steps:
编码：raw text → high-level 写作思路（一句一条）
分析：a. 该思路是否体现了想表达的内容？ b. 思路逻辑是否流畅？
改进：针对不合理处修改思路
解码：改好的思路 → raw text
Why encode and then decode: the surface fluency of raw text hides logical gaps; once compressed into one line per point, logical breaks become conspicuous.
Full example
raw text
Inspired by \tocite{Hier gs, LoG, OctreeGS}, we model scene representation as $\mathcal{G}(v,t,\mu,R,S,o,SH)$ based gaussians primitives. Our method organizes gaussian primitives into a tree-based hierarchy with $L$ levels, higher level leading to coarser visual quality but better efficiency, vice versa. Gaussian primitives in higher level nodes are merged from lower level nodes in a designed way. To be specific, we interpolate all the attributes of level $l$ gaussian primitives … to get attributes of level $l+1$ gaussian primitives. We adaptively select nodes to render according to camera view and manually set nodes' velocity by tracking information.
Encode
1. 用 gaussian primitives 作为 scene representation。
2. 组织为 tree-based hierarchy，higher level 质量更粗但效率更高。
3. Higher level 由 lower level merge 而来。
4. 具体而言，插值 l 层得到 l+1 层。
5. 根据 camera view 渐进式选择 nodes。
Analysis a (does it express what you want to say): what you want to express is ① the specific design of the hierarchy tree ② the motivation for designing it. Diagnosis—① is too rough, it doesn't explain how it is designed; ②is not discussed at all。
Analysis b (is the logic smooth): items 4→5 jump from "how to build the hierarchy" to "how to select nodes for rendering," with no transition.
Improve → decode: add the motivation, refine the design, then rewrite in English.
2. Sentence flow
Definition: the ideas and logic of two adjacent sentences are coherent, with no jumps. Forevery two adjacent sentencesask:
- Does the second sentence continue from the first?
- If it does not continue,is there a transition？
- Does the second sentence introducenew terms, and do they appear abruptly?
3. Is the paragraph clear
Check four points from the reader's perspective: ① Does this paragraph haveone clear theme？ ② first sentencedoes it make clear what this paragraph is about? ③ Everynoun/conceptcan the reader understand it?is it self-contained? ④ Is the logic between adjacent sentences continuous?
Then reverse-outlining: from the paragraphs already written, infer the writing logic behind them and see whether it is coherent. (This is the "encoding" from Section 1—the same action.)
4. Specific to the Method
- Each sentence'smotivationis it clear—the reader must always knowwhy the content in this sentence is being done
- Do the sentences flow
- Are the nouns consistent throughout, don't keep changing them
5. The boundaries of using models as aids
Good at: pointing out sentence flow breaks and logical jumps (the most valuable); turning mixed Chinese-English meaning into idiomatic English; judging whether a collocation is idiomatic. Cannot replace: deciding what this paragraph should express (the level of ideas); judging whether a certain motivation holds.
It is normal to iterate a paragraph's ideas 4–5 times.Revising the ideas in turn while writing specific sentences, is the key to a clear paragraph logic, and is also the step most often skipped.
II. Responding to Reviewer Comments (Rebuttal)
Process
1. 整理 review，回答"为什么 reviewer 给这个分"
2. 分类每条 concern，决定怎么答
3. 起草；标记每个 reviewer 的重点问题，反复确认是否真的答到了
The output of Step 1 is not a list, buta one-sentence explanation of where each reviewer's score comes from— if you cannot answer this, you don't know what to rely on to raise the score.
Classify each concern
| Category | Handling |
|---|---|
| AvailableExisting contentAnswer | Cite specific table/figure/section, do not restate |
| NeededAdd new experiment | First assess whether it can be finished within the rebuttal deadline; if not, downgrade to a limited response |
| YesMisunderstanding | Clarify, and at the same time check whether it isour writingthat caused the misunderstanding |
| It is indeed limitation | Acknowledge, and explain why it does not affect the core contribution |
Language style (the four points that determine success or failure)
- Answer exactly what is asked. Bringing up other things will distract the reviewer.
- Put what the reviewer asked at the very beginning of the paragraph, so that they know immediately what is being answered.
- Answer in the order the reviewer asked the questions as much as possible.
- Under each Reviewer Section, list all of that reviewer's questions and reply to them one by one. Even if there are Common Questions, still write "Refer to Common Questions" under each Section, rather than making the reviewer look elsewhere on their own.
Two bottom lines:Give the reviewer whatever they ask for；Do not introduce new problems in the response(it is very easy to fail because of this).
Key issues
After writing the first draft, markeach reviewer's key issues(usually in the justification), and put most of your energy into confirming these one or two points.
Each reviewer generally has only a few main points of concern. Some reviewers write a lot, but once there is too much they cannot remember it themselves either;they can firmly remember those one or two key points in the review. Of course, the other questions should also be answered well; it is just that the allocation of effort is different.
Output template
## Rebuttal 计划
### 每个 reviewer 的分数从哪来
- R1（分数 X）：<一句话> 重点问题：<1-2 条，引自 justification>
- R2 …
### Concern 分类
| Reviewer | Concern | 类别 | 回应要点 | 依据（table/fig/§ 或需补的实验） | 期限内可行 |
|---|---|---|---|---|---|
### 草稿结构
每个 Reviewer Section 下按其提问顺序逐条；Common Questions 单列并在各 Section 引用。
### 自查
- [ ] 有没有答非所问 - [ ] 有没有引入新问题
- [ ] 重点问题答清楚了 - [ ] 每条回应都指向具体证据