paper-reviewer-simulator

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Paper Reviewer Simulator

论文审稿人模拟器

Run a pre-submission shadow review from the perspective of target-conference reviewers. The goal is to find the objections reviewers are likely to raise before the paper is submitted, then turn those objections into concrete revision priorities.

Use this skill for:

reviewer-style paper critique
venue-specific predicted scores and confidence
likely reject reasons and reviewer questions
adversarial "Reviewer 2" stress tests
area-chair or meta-review summaries
rebuttal-readiness checks
risk register creation for a paper under revision
learning from example reviews, OpenReview discussions, and target-venue guidelines

Do not use this skill to rewrite the paper directly. Pair it with

conference-writing-adapter

after the review if the paper needs structural or paragraph-level changes. Pair it with

citation-audit

for reference correctness and

submit-paper

for final submission hygiene.

Pair this skill with

research-project-memory

when simulated reviewer risks should become project-level risks and actions.

从目标会议审稿人的视角开展提交前的模拟评审。目标是在论文提交前找出审稿人可能提出的异议，进而将这些异议转化为具体的修订优先级。

可将此技能用于：

审稿人风格的论文批评
针对特定会议的预测分数及置信度评估
可能的拒稿原因及审稿人问题梳理
对抗性的「Reviewer 2」压力测试
领域主席或元评审总结
反驳准备情况检查
修订中论文的风险登记册创建
从评审示例、OpenReview讨论及目标会议指南中学习

请勿直接使用此技能重写论文。若论文需要结构或段落层面的修改，可在评审后搭配

conference-writing-adapter

使用；若需检查参考文献正确性，可搭配

citation-audit

；若需完成最终提交的合规检查，可搭配

submit-paper

。

当模拟审稿人提出的风险需要转化为项目级风险及行动时，可搭配

research-project-memory

使用。

Skill Directory Layout

技能目录结构

text

<installed-skill-dir>/
├── SKILL.md
└── references/
    ├── example-review-mining.md
    ├── memory-model.md
    ├── report-template.md
    ├── review-panel.md
    ├── risk-register.md
    └── venue-review-styles.md

text

<installed-skill-dir>/
├── SKILL.md
└── references/
    ├── example-review-mining.md
    ├── memory-model.md
    ├── report-template.md
    ├── review-panel.md
    ├── risk-register.md
    └── venue-review-styles.md

Progressive Loading

渐进式加载规则

Always read

references/review-panel.md

and

references/risk-register.md

Read
```
references/venue-review-styles.md
```
when the user names a target conference or asks to compare venues.
Read
```
references/example-review-mining.md
```
when learning from OpenReview, public reviews, accepted-paper discussions, official reviewer guidelines, or example papers.
Read
```
references/memory-model.md
```
whenever saving or reusing venue-specific reviewer knowledge.
Use
```
references/report-template.md
```
for full review reports.
Verify current official reviewer guidelines and review forms when scoring or venue-specific criteria matter.

始终读取

references/review-panel.md

和

references/risk-register.md

。

当用户指定目标会议或要求对比不同会议时，读取
```
references/venue-review-styles.md
```
。
当需要从OpenReview、公开评审、已录用论文讨论、官方审稿指南或示例论文中学习时，读取
```
references/example-review-mining.md
```
。
当需要保存或复用特定会议的审稿人知识时，读取
```
references/memory-model.md
```
。
使用
```
references/report-template.md
```
生成完整评审报告。
当评分或特定会议标准至关重要时，验证当前官方审稿指南及评审表单。

Core Principles

核心原则

Review like a real reviewer, not like a writing coach. Focus on acceptance risk: novelty, correctness, significance, evidence, clarity, reproducibility, ethics, and fit.
Separate fatal weaknesses from fixable presentation issues.
Ground criticisms in the paper text, figures, experiments, proofs, and citations.
Simulate multiple reviewer personas because one paper can receive conflicting reviews.
Make venue and topic explicit. A theory-heavy ICLR review, empirical NeurIPS review, and CVPR benchmark review should not sound the same.
Do not invent missing results. Mark missing evidence as risk.
Produce actionable fixes, but keep the reviewer's objection distinct from the author's revision plan.

像真实审稿人一样评审，而非写作教练。聚焦录用风险：创新性、正确性、重要性、证据充分性、清晰度、可复现性、伦理合规性及会议适配性。
区分致命缺陷与可修复的表述问题。
基于论文文本、图表、实验、证明及引用提出批评。
模拟多个审稿人角色，因为同一篇论文可能收到相互矛盾的评审意见。
明确会议及主题差异。侧重理论的ICLR评审、侧重实证的NeurIPS评审与CVPR基准测试评审的风格应截然不同。
不得编造缺失的研究结果。将缺失证据标记为风险。
提供可执行的修复方案，但需将审稿人的异议与作者的修订计划区分开。

Step 1 - Define Review Context

步骤1 - 定义评审上下文

Identify:

target venue and year
track or subject area, if known
paper type: method, theory, empirical study, benchmark, dataset, systems, analysis, application, or hybrid
paper stage: outline, early draft, full draft, rebuttal, camera-ready
review mode:
- ```
quick
```
  : top risks and predicted decision
- ```
full
```
  : multi-reviewer reviews plus meta-review and risk register
- ```
adversarial
```
  : skeptical review focused on rejection paths
- ```
rebuttal
```
  : questions and response strategy for an existing review
available evidence: TeX source, PDF, figures, tables, appendix, experiment logs, related work, prior reviews

If the user provides no mode, default to

full

for a complete draft and

quick

for an outline or partial draft.

确定以下信息：

目标会议及年份
赛道或主题领域（若已知）
论文类型：方法类、理论类、实证研究类、基准测试类、数据集类、系统类、分析类、应用类或混合类
论文阶段：大纲、初稿、完整草稿、反驳阶段、终稿
评审模式：
- ```
quick
```
  ：核心风险及预测决策
- ```
full
```
  ：多位审稿人评审+元评审+风险登记册
- ```
adversarial
```
  ：聚焦拒稿路径的质疑式评审
- ```
rebuttal
```
  ：针对已有评审的问题及回应策略
可用证据：TeX源码、PDF、图表、表格、附录、实验日志、相关工作、过往评审意见

若用户未指定模式，完整草稿默认使用

full

模式，大纲或部分草稿默认使用

quick

模式。

Step 2 - Read Paper Evidence

步骤2 - 读取论文证据

Prefer primary paper files over summaries.

Look for:

```
main.tex
```
,
```
paper.tex
```
,
```
sections/*.tex
```
, appendix, supplement
title, abstract, introduction, contribution list
method/theory section, assumptions, theorem statements, algorithm boxes
experiments, baselines, ablations, datasets, metrics, qualitative examples
figures and tables
related work, limitations, ethics/broader impact, checklist
prior reviews or author notes, if available

Build this snapshot:

markdown

undefined

优先读取论文原始文件而非摘要。

重点查看：

```
main.tex
```
、
```
paper.tex
```
、
```
sections/*.tex
```
、附录、补充材料
标题、摘要、引言、贡献列表
方法/理论章节、假设、定理陈述、算法框图
实验、基线模型、消融实验、数据集、指标、定性示例
图表
相关工作、局限性、伦理/更广泛影响、检查清单
过往评审意见或作者说明（若有）

构建如下快照：

markdown

undefined

Paper Snapshot

论文快照

Target venue:
Paper archetype:
Claimed contribution:
Core technical idea:
Main evidence:
Strongest result:
Weakest result:
Most likely novelty concern:
Most likely correctness concern:
Most likely empirical concern:
Current missing information:


If the paper is incomplete, review the current stage honestly and separate "not yet written" from "scientifically weak."

目标会议：
论文类型：
宣称的贡献：
核心技术思路：
主要证据：
最佳研究结果：
最弱研究结果：
最可能的创新性质疑：
最可能的正确性质疑：
最可能的实证性质疑：
当前缺失的信息：


若论文未完成，需如实评审当前阶段，并区分「尚未撰写」与「科学性薄弱」。

Step 3 - Learn Venue and Topic Review Style

步骤3 - 学习会议及主题评审风格

Read

references/venue-review-styles.md

Then update the review context from current sources:

official reviewer guidelines and review form
scoring rubric and confidence scale
ethics/reproducibility/checklist requirements
OpenReview examples and discussions when public
accepted/oral/spotlight papers in the same venue/topic
reviews of similar papers, if accessible

When studying examples, read

references/example-review-mining.md

and produce a compact review-style matrix before scoring the user's paper.

Do not rely on static memory for current review forms. If exact scores or criteria matter, verify them from official sources.

读取

references/venue-review-styles.md

。

随后从当前来源更新评审上下文：

官方审稿指南及评审表单
评分规则及置信度量表
伦理/可复现性/检查清单要求
公开的OpenReview示例及讨论
同一会议/主题下的已录用/口头报告/ spotlight论文
同类论文的评审意见（若可获取）

研究示例时，需读取

references/example-review-mining.md

，并在为用户论文评分前生成简洁的评审风格矩阵。

不得依赖静态记忆存储当前评审表单。若精确评分或标准至关重要，需从官方来源验证。

Step 4 - Configure the Reviewer Panel

步骤4 - 配置审稿人小组

Read

references/review-panel.md

For a full review, create 3-5 reviewers:

```
R1 Technical Specialist
```
: understands the method/theory deeply
```
R2 Skeptical Generalist
```
: asks whether the contribution matters
```
R3 Empirical/Reproducibility Reviewer
```
: checks experiments, baselines, ablations, statistical support
```
R4 Related Work/Novelty Reviewer
```
: checks positioning and missing citations
```
AC Meta-Reviewer
```
: synthesizes decision risk and asks what would change the outcome

Customize the panel to the paper:

theory-heavy paper: add assumptions/proofs reviewer
benchmark/dataset paper: add data quality/evaluation reviewer
systems paper: add scalability/reproducibility reviewer
application paper: add domain-validity reviewer
CV/NLP paper: add task-specific benchmark reviewer

读取

references/review-panel.md

。

对于完整评审，创建3-5位审稿人：

```
R1 技术专家
```
：深入理解方法/理论
```
R2 质疑通才
```
：关注贡献的实际价值
```
R3 实证/可复现性审稿人
```
：检查实验、基线模型、消融实验、统计支持
```
R4 相关工作/创新性审稿人
```
：检查定位及缺失引用
```
AC 元评审人
```
：综合决策风险并提出改变结果的关键因素

根据论文类型定制小组：

侧重理论的论文：增加假设/证明审稿人
基准测试/数据集论文：增加数据质量/评估审稿人
系统类论文：增加可扩展性/可复现性审稿人
应用类论文：增加领域有效性审稿人
CV/NLP论文：增加任务特定基准测试审稿人

Step 5 - Run Independent Reviews

步骤5 - 开展独立评审

Each reviewer should output:

summary of the paper in reviewer voice
strengths
weaknesses
questions for authors
requested experiments/analysis/proofs
score and confidence using the target venue scale when known
likely recommendation: accept / borderline / reject

Reviewers should disagree when reasonable. Do not average away important conflicts.

For each criticism, identify:

location in the paper
what the reviewer is worried about
whether the issue is evidence, clarity, novelty, correctness, positioning, reproducibility, or scope
what would reduce the risk before submission

每位审稿人需输出：

以审稿人视角撰写的论文摘要
优点
缺点
向作者提出的问题
要求补充的实验/分析/证明
已知目标会议评分体系下的分数及置信度
可能的推荐结果：录用 / 边缘 / 拒稿

在合理情况下，审稿人可提出不同意见。不得掩盖重要的意见冲突。

针对每一项批评，需明确：

在论文中的位置
审稿人的担忧点
问题属于证据、清晰度、创新性、正确性、定位、可复现性还是范围类
提交前如何降低该风险

Step 6 - Write Area-Chair Meta-Review

步骤6 - 撰写领域主席元评审

The meta-review should synthesize:

consensus strengths
consensus weaknesses
polarized issues
likely discussion dynamics
decision risk
what would most improve the score
whether the paper is rejected for scientific weakness or presentation weakness

Use this decision language:

```
likely accept
```
```
borderline accept
```
```
borderline reject
```
```
likely reject
```
```
incomplete / not reviewable
```

Add confidence: low / medium / high.

元评审需综合：

共识性优点
共识性缺点
争议性问题
可能的讨论动态
决策风险
最能提升评分的改进方向
论文因科学性薄弱还是表述薄弱而被拒稿

使用以下决策表述：

```
likely accept
```
（大概率录用）
```
borderline accept
```
（边缘录用）
```
borderline reject
```
（边缘拒稿）
```
likely reject
```
（大概率拒稿）
```
incomplete / not reviewable
```
（不完整/无法评审）

添加置信度：低/中/高。

Step 7 - Build Risk Register

步骤7 - 构建风险登记册

Read

references/risk-register.md

Convert reviewer objections into a ranked risk register:

```
must-fix
```
: likely to cause rejection
```
should-fix
```
: materially improves acceptance odds
```
nice-to-fix
```
: polish or reviewer convenience
```
rebuttal-only
```
: cannot fix before submission but can prepare response

Each risk must include:

reviewer concern
evidence from paper
severity
probability
fix effort
recommended action
owner if known

读取

references/risk-register.md

。

将审稿人异议转化为分级风险登记册：

```
must-fix
```
（必须修复）：可能导致拒稿
```
should-fix
```
（建议修复）：显著提升录用概率
```
nice-to-fix
```
（优化项）：润色或提升审稿人体验
```
rebuttal-only
```
（仅需反驳）：提交前无法修复，但可准备回应

每项风险需包含：

审稿人担忧点
论文中的相关证据
严重程度
发生概率
修复工作量
推荐行动
负责人（若已知）

Step 8 - Prepare Rebuttal Readiness

步骤8 - 准备反驳就绪材料

For top risks, produce:

likely reviewer question
best pre-submission fix
fallback rebuttal answer
evidence needed to make the rebuttal credible

If a risk cannot be fixed without new experiments/proofs, say so explicitly.

针对核心风险，生成：

审稿人可能提出的问题
最佳提交前修复方案
备选反驳答案
使反驳具备可信度所需的证据

若某项风险无法通过新实验/证明修复，需明确说明。

Step 9 - Update Reviewer Memory

步骤9 - 更新审稿人记忆

Read

references/memory-model.md

When venue examples or real reviews were studied:

update

.agent/reviewer-simulator/venues/<venue>.md

update

.agent/reviewer-simulator/examples/<venue>-<year>-reviews.md

update

.agent/reviewer-simulator/project-risk-register.md

Memory must separate:

observed reviewer patterns
inferred venue review style
risks for the current paper
unresolved questions

Do not store long copied review text. Paraphrase review patterns and include source URLs.

If the project uses

research-project-memory

, also update:

```
memory/risk-board.md
```
: top simulated reviewer risks, each linked to affected claim IDs when possible
```
memory/action-board.md
```
: must-fix and should-fix actions, distinguishing writing fixes from new experiments/proofs
```
memory/claim-board.md
```
: claims likely to be weakened, cut, or reframed
```
memory/evidence-board.md
```
: evidence gaps, stale figures/tables, or missing proof/experiment needs
```
reviewer/.agent/reviewer-status.md
```
: review mode, predicted decision, and unresolved reviewer questions

Do not treat simulated reviews as real reviews. Use certainty

inferred

for predicted reviewer behavior.

读取

references/memory-model.md

。

当研究了会议示例或真实评审意见时：

更新

.agent/reviewer-simulator/venues/<venue>.md

更新

.agent/reviewer-simulator/examples/<venue>-<year>-reviews.md

更新

.agent/reviewer-simulator/project-risk-register.md

记忆需区分：

观察到的审稿人模式
推断的会议评审风格
当前论文的风险
未解决的问题

不得存储冗长的评审原文。需转述评审模式并包含来源URL。

若项目使用

research-project-memory

，还需更新：

```
memory/risk-board.md
```
：核心模拟审稿人风险，尽可能关联受影响的主张ID
```
memory/action-board.md
```
：必须修复及建议修复的行动，区分写作修改与新实验/证明需求
```
memory/claim-board.md
```
：可能被弱化、删除或重构的主张
```
memory/evidence-board.md
```
：证据缺口、过时图表或缺失的证明/实验需求
```
reviewer/.agent/reviewer-status.md
```
：评审模式、预测决策及未解决的审稿人问题

不得将模拟评审视为真实评审。对预测的审稿人行为使用

inferred

（推断）的确定性标注。

Output Modes

输出模式

Quick Review

快速评审

markdown

undefined

markdown

undefined

Shadow Review: Quick Risk Scan

模拟评审：快速风险扫描

Paper Snapshot

论文快照

Predicted Decision

预测决策

Top 5 Rejection Risks

前5大拒稿风险

Fastest Fixes

最快修复方案

Questions Reviewers Will Ask

审稿人可能提出的问题

undefined

undefined

Full Review

完整评审

Use

references/report-template.md

使用

references/report-template.md

。

Adversarial Review

对抗性评审

Focus on rejection paths:

markdown

undefined

聚焦拒稿路径：

markdown

undefined

Adversarial Review

对抗性评审

Strongest Reject Case

最强拒稿理由

Reviewer 2 Critique

Reviewer 2批评意见

Fatal If True

若属实则致命的问题

How To Disarm Before Submission

提交前如何化解

Risks That Cannot Be Fixed By Writing

无法通过写作修复的风险

undefined

undefined

Rebuttal Mode

反驳模式

markdown

undefined

markdown

undefined

Rebuttal Readiness

反驳就绪材料

Review Claim

评审主张

Is The Reviewer Right?

审稿人是否正确？

Evidence Available

可用证据

Best Response

最佳回应

Paper Revision Needed

需修订的论文内容

undefined

undefined

Final Sanity Check

最终合理性检查

Before finalizing:

target venue/year and review mode are explicit
paper archetype is stated
reviewer criteria come from current official guidance or clearly marked memory
example-review observations are paraphrased and source-linked
criticisms are tied to evidence in the paper
predicted scores include confidence and uncertainty
top risks are ranked by acceptance impact
fixes distinguish writing changes from new experiments/proofs
memory updates are written when venue examples were studied
project memory is updated when simulated risks should persist

定稿前需确认：

目标会议/年份及评审模式明确
论文类型已说明
审稿人标准来自当前官方指南或明确标记的记忆
评审示例观察结果已转述并附带来源链接
批评意见与论文证据相关联
预测分数包含置信度及不确定性
核心风险按录用影响程度排序
修复方案区分写作修改与新实验/证明需求
研究会议示例时已更新记忆
当模拟风险需持续跟踪时已更新项目记忆