conference-rebuttal

v1.0.0

Use when writing or planning rebuttals for peer-reviewed conference papers (ICML, NeurIPS, ICLR, CVPR, etc.).

0· 103·0 current·0 all-time
byZHOU PENGFEI@lancezpf

Install

OpenClaw Prompt Flow

Install with OpenClaw

Best for remote or guided setup. Copy the exact prompt, then paste it into OpenClaw for lancezpf/conference-rebuttal.

Previewing Install & Setup.
Prompt PreviewInstall & Setup
Install the skill "conference-rebuttal" (lancezpf/conference-rebuttal) from ClawHub.
Skill page: https://clawhub.ai/lancezpf/conference-rebuttal
Keep the work scoped to this skill only.
After install, inspect the skill metadata and help me finish setup.
Use only the metadata you can verify from ClawHub; do not invent missing requirements.
Ask before making any broader environment changes.

Command Line

CLI Commands

Use the direct CLI path if you want to install manually and keep every step visible.

OpenClaw CLI

Bare skill slug

openclaw skills install conference-rebuttal

ClawHub CLI

Package manager switcher

npx clawhub@latest install conference-rebuttal
Security Scan
VirusTotalVirusTotal
Benign
View report →
OpenClawOpenClaw
Benign
high confidence
Purpose & Capability
Name/description (conference rebuttal help) match the SKILL.md. The skill is instruction-only and asks for no binaries, env vars, installs, or external services — all proportional to a writing/strategy guide.
Instruction Scope
SKILL.md contains detailed, actionable guidance and an agent workflow (analyze reviews, draft replies, compress to page limits). It suggests running supplemental experiments early, but it does not supply commands or request system access to run them — this is guidance for users/authors, not an automated experiment runner. The skill expects the user to provide review text and (potentially) paper excerpts, which is sensible but raises confidentiality considerations.
Install Mechanism
No install spec and no code files — lowest risk. Nothing will be written to disk or downloaded as part of the skill itself.
Credentials
Requires no environment variables, credentials, or config paths. The only implicit requirement is that the user supplies review content and possibly parts of the manuscript; that is a data-privacy concern rather than an overbroad credential request.
Persistence & Privilege
always is false and the skill does not request special privileges. disable-model-invocation is false (default) which is normal; autonomous invocation is allowed by platform defaults but the skill itself does not demand permanent presence or modify other skills.
Assessment
This skill is a coherent, instruction-only guide for drafting rebuttals and does not ask for credentials or install anything. Before using it, consider: (1) confidentiality — you will likely paste reviewer comments and unpublished manuscript text into the agent; avoid sharing full sensitive data if you need to keep IP private, or redact identifying details; (2) experimental claims — the guide recommends running supplemental experiments early, but the agent cannot run experiments for you; do not let the agent fabricate numbers or results — always verify any quantitative claims it produces; (3) autonomy — the skill can be invoked autonomously by the agent under normal platform defaults; if you prefer to control when the skill runs, restrict autonomous invocation in your agent settings; (4) verification — use the generated rebuttal as a draft to edit and verify references/line numbers and to ensure nothing misrepresents the paper. If you want the agent to run computations or access private datasets, expect to need additional tooling/credentials not provided by this skill.

Like a lobster shell, security has layers — review code before you run it.

latestvk97f6d7evw0cak4ap5g6xxbnkd8412z8
103downloads
0stars
1versions
Updated 3w ago
v1.0.0
MIT-0

顶会 Rebuttal 写作指南

核心原则

Rebuttal 的唯一战略目标:武装你的 Champion(Arm the Champion)

Champion 是愿意为你的论文辩护的审稿人。Rebuttal 是给 Champion 提供弹药,让 TA 在讨论阶段替你据理力争。AC 很可能只读 review + rebuttal 而不重读论文,所以 rebuttal 必须自包含弱 rebuttal 会让 Champion 放弃支持并降分。

心法:永远不是在"辩论",而是在proactively"展示"。 你的姿态始终是"感谢您的问题,让我展示更多数据来澄清",而不是"您的质疑是错误的"。

受众策略

受众分类目标投入
ChampionDefend巩固支持,提供弹药认真回应所有意见,不可敷衍
摇摆审稿人Flip(核心目标)消除疑虑,争取翻转最大深度,补充实验,主动检索
强反对者Neutralize纠正事实错误,降低影响力聚焦误解澄清,不过度投入
AC帮 TA 快速判断所有意见是否被合理回应AC 速览信

四层架构

1. Overall      — 致全体的总结信(借审稿人之口夸自己,有时候没有这个按钮,optional)
2. To AC        — 一页纸速览(每个问题压缩成一句话回答)
3. To Reviewer X — 逐人逐条回复(W1, W2... Q1, Q2... 编号与原 review 一一对应)
4. To AC (终稿)  — 如遇不公审稿,最后补一封事实驱动的申诉信

排序原则:从最友善的审稿人开始,最有争议的放最后。让 AC 先建立"问题都能解决"的信心。

流程

  1. 逐条拆解 — 表格整理每条意见,标注:类别(误解/实验/写作/真缺陷)、严重程度、是否多人共同提出
  2. 识别 Champion — 评分最高或评语最积极的审稿人,TA 的意见优先且认真回应
  3. Brain Dump → 凝练 — 每位作者先独立写粗糙回复,再合并精简。好 rebuttal 是删出来的。
  4. 立即启动补充实验 — 时间是最稀缺资源,尽早跑

写作策略

结构层

  • 借嘴说话(Opening 黄金公式) — 从每位审稿人的 review 中各挑一句最正面的原话加粗引用。即使给低分的审稿人也能找到正面词。TA 自己说的好话,后续攻击的分量自动被平衡
    We are encouraged that R1 acknowledges "**well motivated**",
    R2 finds it "**reasonable and clever**", R3 describes it as
    "**conceptually elegant and practically impactful**".
    
  • AC 速览信 — 用嵌套列表把所有问题压缩成一屏摘要,关键词加粗,让 AC 扫一眼就能判断"已解决"
  • 合并同类 — 多位审稿人的相同问题合并回答,节省篇幅

回应层(单条五步公式)

① > 引用审稿人原文(blockquote)
② 感谢语(一句话,变换措辞避免重复)
③ 直接结论(Yes/No/Not quite,加粗核心论断)
④ 数据/表格/实验支撑
⑤ "已更新至论文 Section X / Lines Y-Z / Appendix M"

第⑤步是隐藏武器——信号:我们不只是嘴上回应,论文已经改了。

  • 回应意图而非字面 — "为何不在数据集 Z 上测?" 真实意图是质疑实验充分性。回答 Z 后主动提醒你已在 X/Y/W 上做了充分评估
  • 做了再说,别光许诺 — 在 rebuttal 里直接展示内容,再补一句"已加入终稿"

证据层

  • 数据 > 论证 > 承诺 — 每次发现自己在"说服",问:能不能用一个数字替代这段话?
  • 精确量化 — 不说"开销很小",说 "overhead is only 2.2ms/step (<1% latency)"。技巧:"only"+绝对值、"less than X%"、"comparable to [已接受方法]"
  • 超额完成 — 审稿人要一个 benchmark?不止给总分,还拆 5 个维度全面碾压。传递信号:经得起任何角度审视
  • Claim credit — 论文里已有的内容,给行号/表号,然后在 rebuttal 中重述。行号建立可信度,重述让 AC 不用翻论文
  • 重构已有证据 — 论文里有相关数据但没有从这个角度解读?重新框架现有结果来回答新问题(零成本防守)

高级策略

  • 交叉引用审稿人(结盟术) — R1 质疑某方面,R2/R3 恰好肯定了它?引用:"as also highlighted by R2 and R3",制造 2 对 1 局面
  • 正交性论证(化敌为友) — 被要求与竞品比较且不占绝对优势?论证互补性:"highlights the orthogonal nature and potential complementarity"
  • 贡献差异化(回应 novelty 质疑) — ① 承认灵感来源(诚实)→ ② 指出先前工作的具体局限 → ③ 用实验数据证明差异有效果 → ④ 引用审稿人自己在 Strengths 中说的正面评价来反驳
  • 反引审稿人自己的话 — 审稿人 Strengths 里的正面评价与 Weaknesses 矛盾时,直接引用回去。心理上很难反驳自己说过的话

语气层

  • 对话式而非对抗式 — 建设性学术讨论,不是辩论赛
  • 关键问题有深度 — 不是 yes/no,需多角度:证据/普遍挑战/专业洞察/坦诚局限
  • 接纳合理建议 — 审稿人建议换 baseline?直接跑。结果好双赢,结果差也是有力论据
  • 感谢用心审稿 — 列了 typo?提供文献?表达感谢。同行不是敌人

对付不公审稿(阶梯式升级)

技术回复中始终礼貌,但在终稿 AC 信中可以直接有力:

Round 1:耐心解释 + 数据 → Round 2:更坚定纠正 + 文献支撑 → To AC 信:结构化投诉

投诉信结构:① 陈述审稿人的具体不合理要求 → ② 量化不合理性("需 6,811 A800 GPU hours")→ ③ 引用领域通行做法 → ④ 请求 AC 公正裁决

Discussion 阶段

Rebuttal 提交后不是结束:

  • 持续补充跑完的新实验结果
  • Deadline 前主动 follow up:"We would like to kindly follow up..."
  • 为长期未回复的审稿人提供精简总结版

排版要点

元素用法目的
加粗只加粗结论性关键词和关键数字扫读者抓住要点
> 引用块完整引用审稿人原文上下文清晰可查
表格每个实验/分析一张表数据比文字有说服力
Conclusion 段每张表后一句话加粗总结不看表也能懂
Section 标注"已加入 Appendix M"证明修改已落地
嵌套列表多层论证用缩进逻辑层次一目了然

常见错误

错误正确做法
只回复反对者,忽视 ChampionChampion 的意见最优先回应
"详见论文第 X 节"在 rebuttal 中重述,行号仅作辅助
"我们会在终稿中加入..."在 rebuttal 中直接展示内容
逐字反驳每个小问题合并同类、抓大放小
情绪化、对抗式语气对话式、用数据说话
回避真实缺陷坦诚承认 + 缓解方案 + 未来工作
模糊定性描述精确量化(only 2.2ms, <1%)
虚构证据造假核对论文原文中真的是否有refer的信息,引用的论证是否真实存在

Agent 工作流

当用户提供 review 内容后:

  1. 分析 — 逐条拆解,分类为:误解、实验不足、写作问题、真实缺陷、建设性建议
  2. 策略 — 识别 Champion,分类审稿人(Defend/Flip/Neutralize),确定优先级
  3. 起草 — 按四层架构 + 五步公式生成 rebuttal 草稿,确保每条意见都被覆盖
  4. 精简 — 删减至字数限制,优先保留数据支撑的回应,压缩纯文字论证
  5. 自检 — 不看原文的人能否仅凭 rebuttal 理解所有回应?

Comments

Loading comments...