Overview
This skill provides a systematic academic paper review process, suitable for peer review at top conferences (such as NeurIPS, ICML, ICLR, ACL, CVPR, etc.) and journals. The entire process is divided into four steps. The first three steps conduct in-depth analysis in Traditional Chinese, and the fourth step produces a formal English review that conforms to conference formatting.
Core Principles
- Rigorous but fair: The purpose of review is to improve academic quality, not to suppress authors
- Specific rather than vague: All criticisms should point to specific sections, equations, or experiments
- Constructive orientation: While pointing out problems, provide directions for improvement
- Multidimensional evaluation: Do not judge by a single standard; comprehensively consider all aspects of the paper
Four-Step Review Process
Overview
| Step | Name | Language | Purpose |
|---|---|---|---|
| Step 1 | Critical review | Traditional Chinese | Read in depth and identify strengths and weaknesses |
| Step 2 | Score prediction | Traditional Chinese | Scoring by dimension and overall score prediction |
| Step 3 | Key point refinement | Traditional Chinese | Organize findings into structured key points |
| Step 4 | Formal review | English | Produce a formal review that conforms to conference formatting |
Step 1: Critical Review (Traditional Chinese)
Objective
Read the paper in depth with critical thinking and systematically identify strengths and weaknesses. This step is the foundation for all subsequent analysis.
Execution Method
-
First pass: quick skim (5-minute mindset)
- Read the title, abstract, introduction, and conclusion
- Skim figures and tables
- Form an initial impression: What is this paper doing? What contributions does it claim?
-
Second pass: careful reading (30-minute mindset)
- Read section by section and mark unclear places
- Check the correctness of mathematical derivations
- Verify the reasonableness of the experimental setup
- Confirm whether the claims are sufficiently supported
-
Third pass: critical analysis (15-minute mindset)
- Evaluate one by one against the five major review dimensions
- Identify whether it matches the five major rejection patterns
- Think about missing baselines, missing ablations, missing analysis
Output Format
## Step 1：批判性審查
### 論文基本資訊
- 標題：
- 作者：
- 投稿會議/期刊：
- 領域：
### 初步印象
（一段話總結論文做了什麼、聲稱了什麼）
### 優點列表
1. [S1] （具體描述優點，引用論文段落或頁碼）
2. [S2] ...
### 缺點列表
1. [W1] （具體描述缺點，說明為什麼這是問題）
2. [W2] ...
### 疑問列表
1. [Q1] （需要作者釐清的問題）
2. [Q2] ...
### 初步判斷
（對論文品質的初步判斷，包含是否達到會議門檻的直覺）
Notes
- Both strengths and weaknesses must be specific; avoid vague comments such as "well written" or "insufficient experiments"
- Each strength and weakness should correspond to a specific location in the paper
- The question list should distinguish between "genuine questions" and "rhetorical criticisms"
- The initial judgment is only intuition; do not be anchored by it — subsequent steps may change the conclusion
Step 2: Score Prediction (Traditional Chinese)
Goal
Based on the analysis in Step 1, score the paper across five dimensions and predict the overall score.
Five Review Dimensions
Each dimension is scored on a 1-5 scale. For detailed scoring criteria, see references/scoring-rubric.md。
| Dimension | Description | Weight Reference |
|---|---|---|
| Novelty & Contribution | Is the idea novel? Is the contribution sufficient? | High |
| Clarity & Writing | Is the paper easy to read? Is the logic clear? | Medium |
| Experimental Rigor | Is the experimental design rigorous? Are the results reliable? | High |
| Evaluation Completeness | Is the evaluation comprehensive? Are there missing comparisons? | Medium-High |
| Soundness of Method | Is the method correct? Are the assumptions reasonable? | High |
For detailed dimension descriptions, please refer to references/review-dimensions.md。
Scoring System
- Overall Score: 1-10 points, see
references/scoring-rubric.md - Confidence: 1-5 points, reflecting the reviewer's confidence in their own assessment
- Scores for each dimension: 1-5 points
Output Format
## Step 2：分數預測
### 各維度評分
| 維度 | 分數 (1-5) | 簡要理由 |
|------|-----------|----------|
| 貢獻新穎性 | X | （一句話理由） |
| 寫作清晰度 | X | （一句話理由） |
| 實驗嚴謹性 | X | （一句話理由） |
| 評估完整性 | X | （一句話理由） |
| 方法健全性 | X | （一句話理由） |
### 整體分數預測
- **Overall Score**: X / 10
- **Confidence**: X / 5
### 分數推理
（解釋為什麼給這個整體分數，各維度如何影響最終判斷）
### 與接受門檻的比較
（討論這篇論文相對於目標會議的接受門檻如何，是 clear accept、borderline 還是 clear reject）
Notes
- Scores for each dimension should be consistent with the strengths and weaknesses in Step 1
- The overall score is not a simple average of the dimension scores—some fatal flaws will directly lower the overall score
- Confidence should be honest—if you are not familiar with this subfield, your confidence should be lower
- Consider the conference's acceptance rate (e.g., NeurIPS ~25%, ICML ~25%, workshop ~50%)
Step 3: Refining Key Points (Traditional Chinese)
Goal
Organize the analyses from Step 1 and Step 2 into refined structured key points, serving as a draft for the formal review in Step 4.
Execution Method
- Merge and deduplicate: Merge the strengths and weaknesses from Step 1, removing duplicates
- Sort: Sort by importance, placing the most important first
- Add details: Add specific paper citations and suggestions for each key point
- Classify: Distinguish between major concerns and minor concerns
- Write suggestions: Provide specific improvement suggestions for each weakness
Output Format
## Step 3：要點精煉
### 核心優點（按重要性排序）
1. [Major S1] （精煉描述 + 論文引用）
2. [Major S2] ...
### 重大問題（Major Concerns）
1. [Major W1] （精煉描述 + 為什麼重要 + 建議改進方向）
2. [Major W2] ...
### 次要問題（Minor Concerns）
1. [Minor W1] （精煉描述 + 建議改進方向）
2. [Minor W2] ...
### 必問問題（Questions for Authors）
1. [Q1] （精煉問題 + 為什麼這個答案很重要）
2. [Q2] ...
### 一句話總結
（用一句話總結對這篇論文的評價）
Notes
- The distinction between Major and Minor should be clear—Major refers to issues that affect whether the paper is accepted
- Each weakness should be accompanied by a suggestion, embodying the spirit of constructive reviewing
- Questions should be strategic—good questions can help you make better judgments after the rebuttal
- For how to write constructive feedback, please refer to
references/constructive-feedback.md
Step 4: Formal Review (English)
Goal
Based on the Traditional Chinese analysis from the first three steps, produce a formal English review that conforms to the conference format.
Execution Method
- Use
templates/review-form.mdthe template in - Translate and polish the key points from Step 3 into academic English
- Ensure the tone is professional, objective, and constructive
- Check whether any important strengths or weaknesses have been overlooked
Formal Review Structure
## Summary
（2-4 句話總結論文的主要貢獻與方法）
## Strengths
1. [S1] ...
2. [S2] ...
## Weaknesses
1. [W1] ...
2. [W2] ...
## Questions for Authors
1. [Q1] ...
2. [Q2] ...
## Suggestions for Improvement
1. ...
2. ...
## Minor Issues
- （格式、typo、文法等小問題）
## Rating
- Overall: X/10
- Confidence: X/5
## Justification
（簡短解釋你的評分理由）
Tone Guidelines
- Avoid：「This paper is bad.」「The authors clearly don't understand...」
- Recommend：「The paper would benefit from...」「It is unclear how...」「A potential concern is...」
- Use conventional academic review phrasing (refer to
references/constructive-feedback.md） - Criticize the work, not the person—review the paper, not the authors
Notes
- The Summary should accurately reflect the paper's content so the authors know you actually read the paper
- Strengths should be sincere—even if you are going to reject the paper, acknowledge its merits
- Weaknesses should be specific—point out specific sections, equations, experiments
- Questions should be strategic—they should help you make a decision after the rebuttal
- For reference on rejection patterns, see
references/rejection-patterns.md
Five Review Dimensions (Detailed Version)
The following is a summary of the five review dimensions. For the full explanation, please refer to references/review-dimensions.md。
1. Novelty & Contribution
- Has this idea been done before?
- Is the quantity and quality of the contribution sufficient to support a full paper?
- How large is the increment relative to prior work?
- Is it technically novel, novel in application, or novel in perspective?
2. Clarity & Writing
- Can a non-expert understand the paper's main contributions?
- Is the motivation clear? Why is this problem important?
- Are the notation and terminology consistent?
- Are the figures and tables clear and helpful for understanding?
3. Experimental Rigor
- Is the experimental design sufficient to validate the claims?
- Are there appropriate baselines?
- Are statistical significance or error bars reported?
- Is the hyperparameter selection reasonable? Was a fair comparison conducted?
4. Evaluation Completeness
- Are the datasets sufficiently diverse?
- Is there an ablation study?
- Is there analysis (error analysis, case study)?
- Are the limitations discussed?
5. Soundness of Method
- Is the mathematical derivation correct?
- Are the assumptions reasonable and explicitly stated?
- Does the method work as claimed?
- Is there a theoretical guarantee or intuitive explanation?
Five Rejection Patterns
The following is a summary of the five rejection patterns. For the full explanation and examples, please refer to references/rejection-patterns.md。
Pattern 1: Insufficient Contribution
- Symptoms: incremental improvement, trivial extension, simple combination of known results
- Typical trigger phrase: "This is a straightforward extension of [X]."
Pattern 2: Lack of Clarity
- Symptoms: hard to understand, logical leaps, confusing notation, lack of intuition
- Typical trigger phrase: "The paper is difficult to follow."
Pattern 3: Weak Results
- Symptoms: insignificant improvements, cherry-picked results, lack of statistical tests
- Typical trigger phrase: "The improvements are marginal and may not be statistically significant."
Pattern 4: Incomplete Evaluation
- Symptoms: missing important baselines, too few datasets, no ablation
- Typical trigger phrase: "Key baselines such as [X] are missing."
Mode 5: Flawed Methodology
- Symptoms: mathematical errors, unreasonable assumptions, data leakage, unfair comparisons
- Typical trigger phrase: "There appears to be an error in the derivation of Eq. (X)."
Scoring System Quick Reference
Overall Score (1-10)
| Score | Meaning | Recommendation |
|---|---|---|
| 10 | Top-tier paper, can be accepted directly | Strong Accept |
| 8-9 | Excellent paper, clearly above the threshold | Accept |
| 6-7 | Good paper, slightly above or near the threshold | Weak Accept / Borderline |
| 5 | At the threshold, with clear strengths and weaknesses | Borderline |
| 3-4 | Below the threshold, with major issues | Weak Reject / Reject |
| 1-2 | Clearly not qualified | Strong Reject |
Confidence (1-5)
| Score | Meaning |
|---|---|
| 5 | Very confident — I am an expert in this subfield |
| 4 | Confident — I am familiar with this field |
| 3 | Moderate — I understand the main concepts but am not an expert |
| 2 | Not very confident — some content is outside my expertise |
| 1 | Not confident — this is not my field |
For the complete scoring criteria, please refer to references/scoring-rubric.md。
References and Templates
References
references/review-dimensions.md— Complete explanation of the five review dimensionsreferences/scoring-rubric.md— Complete scoring criteriareferences/rejection-patterns.md— Detailed explanation and examples of the five rejection modesreferences/constructive-feedback.md— Guide to writing constructive feedback
Templates
templates/review-form.md— Standard review form templatetemplates/meta-review-form.md— AC meta-review template
Related Skills
- paper-writing skill — paper writing skill, can be used together for a revision cycle
How to Use
Basic Usage
請幫我審查這篇論文：[貼上論文內容或提供 PDF 路徑]
Execute the four steps in order, ultimately producing a formal English review.
Specify Steps
請只執行 Step 1（批判性審查）
請從 Step 3 開始（我已經有前兩步的分析）
Specify Conference
請以 NeurIPS 2026 的標準審查這篇論文
請以 ACL 2026 的標準審查，特別注意 reproducibility
Meta-review
以下是三位 reviewer 的意見，請幫我撰寫 meta-review：
[貼上三份 review]
Use the templates/meta-review-form.md template.
Review Ethics Reminders
- Confidentiality: Review content must not be disclosed to third parties
- Conflict of Interest: If you know the authors or have a conflict of interest with the paper, you should proactively recuse yourself
- Respect: Even if the paper quality is poor, the review should be written with a respectful attitude
- Timeliness: Completing the review before the deadline is the reviewer's responsibility
- Consistency: Your score should be consistent with your written comments — do not give a high score while writing a full list of shortcomings
- Constructiveness: The ultimate purpose of review is to help the authors improve the paper, not to show off one's own knowledge
Frequently Asked Questions
Q: What should I do if I am not sure whether a certain weakness is correct?
Write it in Questions for Authors and let the authors respond in the rebuttal. Uncertain criticisms should not be used as the main reason for rejection.
Q: How should borderline papers be handled?
Carefully weigh the relative importance of the strengths and weaknesses. If the strengths are clear and significant, and the weaknesses can be fixed in the camera-ready version, lean toward acceptance. If the core method has problems, lean toward rejection.
Q: What is the difference between a meta-review and an individual review?
A meta-review needs to synthesize all reviewers' opinions, identify consensus and disagreement, and make a final recommendation. The AC needs to judge which reviewers' opinions are more valuable and whether the rebuttal effectively addresses the main issues.
Q: How can it be used together with the paper-writing skill?
After completing the review, the suggestions in the review can be passed to the paper-writing skill to guide the author in making revisions. This forms a "review -> revise -> re-review" cycle.