paper-review

name	paper-review
description	AI 顶会论文审稿知识库。包含 ICLR/ICML/NeurIPS/CVPR/ECCV/ICCV 的审稿标准、评分体系、常见表达和伦理准则。用于辅助审稿人撰写专业、建设性的审稿意见。

Paper Review Skill

专业论文审稿知识库，辅助审稿人撰写高质量审稿意见。

重要: 本 skill 仅用于辅助审稿人组织思路，所有审稿意见必须基于审稿人的独立判断，严禁直接提交 AI 生成内容。

---

各会议评分标准

ICLR (International Conference on Learning Representations)

Rating Scale (1-10):

Score	Label	描述
10	Award quality	Top 2%，开创性工作
8	Strong Accept	Top 10%，优秀贡献
6	Weak Accept	高于门槛，可以接收
5	Borderline	边缘，优缺点相当
3	Weak Reject	低于门槛，需要改进
1	Strong Reject	明显问题，不适合该会议

Confidence Scale (1-5):

Score	描述
5	该领域专家，非常熟悉相关工作
4	熟悉该领域，理解论文细节
3	了解大部分内容，但有些部分不确定
2	对该领域了解有限
1	几乎不了解该领域

特点: 公开审稿，作者可回复，审稿人需参与讨论

---

ICML (International Conference on Machine Learning)

Rating Scale:

Score	Label	描述
10	Top 5%	必须接收
9	Top 15%	强接收
7-8	Top 30%	接收
6	Top 50%	边缘接收
5	Borderline	边缘拒绝
4	Reject	拒绝
1-3	Strong Reject	强拒

特点: 双盲审稿，Area Chair 协调

---

NeurIPS (Neural Information Processing Systems)

Rating Scale (1-10):

Score	描述
9-10	Top 2%，突破性工作
7-8	Strong Accept，优秀
6	Weak Accept，可接收
5	Borderline，边缘
4	Weak Reject，略低于门槛
1-3	Strong Reject，明显问题

特点:

• 双盲审稿
• 要求 Broader Impact 和 Limitations 章节
• Author response 阶段

---

CVPR/ICCV/ECCV (计算机视觉顶会)

Rating Scale (CVPR/ICCV):

Score	Label	描述
5	Strong Accept	优秀，应该接收
4	Accept	可以接收
3	Borderline	边缘
2	Reject	应该拒绝
1	Strong Reject	强烈拒绝

ECCV 特点:

• 使用 CMT 系统
• 两轮审稿 (部分投稿)
• 欧洲时区

---

评审维度详解

1. Novelty（新颖性）

评估问题:

• 核心 idea 是否新颖？
• 与现有工作的区别是否显著？
• 是否仅是现有方法的简单组合？

等级标准:

等级	描述	表达方式
高	全新问题定义或方法范式	"first to...", "novel paradigm", "fundamentally new"
中	显著改进现有方法	"interesting extension", "non-trivial improvement"
低	增量式改进	"incremental", "straightforward extension"
无	已有方法	"essentially the same as", "lacks novelty"

常见表达:

高新颖性:
- "This paper introduces a fundamentally new perspective on..."
- "To the best of my knowledge, this is the first work to..."
- "The proposed formulation opens up new research directions."

中等新颖性:
- "While the individual components are known, their combination is novel."
- "The paper provides interesting insights into..."
- "The extension to [new setting] is non-trivial."

低新颖性:
- "The approach is similar to [prior work], with the main difference being..."
- "The contribution seems incremental over [citation]."
- "I'm not convinced the technical novelty is sufficient for [venue]."

---

2. Technical Soundness（技术正确性）

评估问题:

• 方法是否正确？证明是否严谨？
• 假设是否合理？是否明确说明？
• 实验设计是否正确？

常见问题类型:

问题类型	示例	严重程度
证明错误	定理证明有漏洞	Major
假设过强	假设在实际中不成立	Major
实验设置不当	数据泄露、不公平对比	Major
细节缺失	关键参数未说明	Minor
符号混乱	符号定义不一致	Minor

表达方式:

发现问题:
- "I believe there is an issue in the proof of Theorem 1..."
- "The assumption in Section 3.2 seems too strong because..."
- "The experimental setup may have potential data leakage..."

不确定:
- "I may have missed something, but it seems like..."
- "Could the authors clarify how [X] is handled?"
- "Unless I misunderstand, the derivation in Eq. (5) assumes..."

---

3. Empirical Evaluation（实验评估）

评估问题:

• Baseline 是否合理且足够？
• 数据集是否有代表性？
• 消融实验是否充分？
• 结果是否统计显著？

Baseline 评估标准:

等级	描述
优	包含最新 SOTA 方法，公平对比
良	包含主要 baseline，但缺少部分最新工作
中	Baseline 较弱或过时
差	明显缺失重要 baseline

常见表达:

Baseline 问题:
- "The recent work [X] addresses a very similar problem but is not compared."
- "Some baselines are from 2019, and more recent methods should be included."
- "It would be helpful to compare with [specific method] which is SOTA on [dataset]."

消融不足:
- "The ablation study is incomplete. Specifically, the effect of [X] is not studied."
- "It's unclear which component contributes most to the improvement."

统计显著性:
- "Are the improvements statistically significant? Error bars would be helpful."
- "The variance across runs should be reported."

---

4. Presentation（论文呈现）

评估问题:

• 写作是否清晰流畅？
• 图表是否信息丰富？
• 组织结构是否合理？

常见表达:

正面:
- "The paper is well-written and easy to follow."
- "Figure 2 provides a clear overview of the method."
- "The notation is consistent throughout."

负面:
- "The paper could benefit from better organization."
- "Some parts are hard to follow, particularly Section 3.2."
- "The notation is inconsistent between Sections 2 and 3."

建议:
- "A figure illustrating the pipeline would help readers understand the method."
- "Consider moving [X] to the appendix to make room for [Y]."

---

5. Significance（重要性）

评估问题:

• 解决的问题是否重要？
• 方法是否有广泛适用性？
• 对领域发展是否有推动作用？

常见表达:

高重要性:
- "This addresses a fundamental problem in..."
- "The results could have significant impact on [application]."
- "I expect this work to inspire follow-up research."

中等重要性:
- "While the problem is somewhat niche, the solution is elegant."
- "The contribution is solid, though the scope is limited."

低重要性:
- "The practical impact seems limited given [constraint]."
- "It's unclear who would benefit from this method."

---

自然语言表达库

开场白（Summary 开头）

✓ "This paper proposes..."
✓ "This work presents..."
✓ "The authors introduce..."
✓ "This submission addresses..."
✓ "The paper tackles the problem of..."

✗ "This amazing paper..." (过于热情)
✗ "The authors claim..." (暗示不信任)
✗ "This paper attempts to..." (暗示失败)

表达优点

强肯定:
- "A key strength of this work is..."
- "I particularly appreciate..."
- "The most compelling aspect is..."
- "This is a significant contribution because..."

中等肯定:
- "The paper does a good job of..."
- "I found [X] to be interesting."
- "The approach is reasonable and well-motivated."

弱肯定:
- "On the positive side,..."
- "One merit of the paper is..."

表达缺点

委婉批评:
- "One area that could be improved is..."
- "A potential concern is..."
- "I have some reservations about..."
- "One aspect that deserves more attention is..."

中等批评:
- "The paper falls short in..."
- "A weakness of the current version is..."
- "I'm not fully convinced that..."

直接批评（仅用于严重问题）:
- "There appears to be an error in..."
- "The experimental setup is flawed because..."
- "This claim is not supported by the evidence."

表达不确定

- "I may have missed something, but..."
- "Unless I misunderstand,..."
- "Correct me if I'm wrong, but..."
- "I'm not entirely sure about [X], so I'd appreciate clarification."
- "This might be addressed in the appendix, but I couldn't find..."

提出建议

- "It would be helpful if..."
- "The paper would benefit from..."
- "I encourage the authors to..."
- "Consider adding..."
- "One way to strengthen the paper is to..."

评分陈述

弱接收:
- "Overall, I lean towards acceptance, but [concern] should be addressed."
- "The contribution is solid, and I'm inclined to accept."
- "Despite some weaknesses, I believe this paper makes a contribution worth publishing."

边缘:
- "I'm on the fence about this paper."
- "The strengths and weaknesses roughly balance out."
- "My assessment could change depending on the authors' response."

弱拒绝:
- "While the paper has merits, the issues outweigh them."
- "In its current form, I cannot recommend acceptance."
- "The paper needs significant revision before it's ready for publication."

---

避免 AI 检测的写作技巧

1. 使用个人化表达

✓ "In my experience with similar methods,..."
✓ "When I ran into a similar problem in my own work,..."
✓ "I found myself confused by..."
✓ "Honestly, I'm not sure about..."

2. 加入具体细节和位置

✓ "In Line 235, the notation changes from x to y without explanation."
✓ "Table 3, row 4 shows an unexpected result..."
✓ "The derivation between Eq. (5) and (6) skips several steps."

3. 使用口语化连接词

✓ "That said,..."
✓ "Having said that,..."
✓ "On a related note,..."
✓ "Speaking of which,..."

4. 表达情感和反应

✓ "I was surprised to see..."
✓ "I found it interesting that..."
✓ "This is a bit disappointing because..."
✓ "I'm curious about..."

5. 承认自己的局限

✓ "I'm not an expert in [specific area], so take my comment with a grain of salt."
✓ "I might be wrong here, but..."
✓ "This is outside my main area of expertise, but..."

6. 不要过于完美

✓ 适当使用缩写: "doesn't", "I'd", "won't"
✓ 偶尔用非正式表达: "a bit", "kind of", "sort of"
✓ 避免过于对称的结构

---

回复 Rebuttal 的技巧

当问题被解决

"The authors have addressed my concerns regarding [X]. I'm raising my score to [Y]."
"The additional experiments in the rebuttal are convincing. My main concern is resolved."

当问题部分解决

"I appreciate the authors' response. While [X] is addressed, I still have concerns about [Y]."
"The clarification helps, but I'm not fully convinced because..."

当问题未解决

"Unfortunately, the rebuttal does not address my main concern about [X]."
"I appreciate the effort, but the response doesn't change my assessment because..."

---

伦理检查清单

审稿前确认：

• 我没有与作者的利益冲突
• 我有足够的专业知识评审这篇论文
• 我会保护论文内容的机密性
• 我会给出建设性的反馈

审稿中确认：

• 我的所有批评都基于论文内容，没有编造
• 我使用了专业、尊重的语言
• 我对所有论文使用了一致的标准
• 我考虑了论文的优点和缺点

审稿后确认：

• 我没有泄露论文内容
• 我参与了 rebuttal 讨论（如需要）
• 我的最终评分反映了我的诚实判断

---

References

• ICLR 2024 Reviewer Guidelines
• NeurIPS 2024 Reviewer Tutorial
• How to Write Good Reviews (CVPR)
• The Art of Reviewing (Yoshua Bengio)