本文へ移動
cccskills
無料GitHub で公開

citation-verification

Verification guide for citations added during rebuttal writing. Use when Stage 2 responses introduce new references, when the Area Chair asks about a cited paper, or when any citation in the rebuttal might have been AI-generated. Prevents the serious credibility damage of fabricated references in reviewer-facing documents.

インストール方法を見る

含まれるファイル(1)

  • SKILL.md5.8 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

Citation Verification for Rebuttals

RebuttalStudio Utility — Stage 2 / Stage 4 Apply this skill whenever a Stage 2 draft or Stage 4 follow-up response introduces a new reference — especially when the reference was suggested by an LLM. A single fabricated citation in a rebuttal can undermine the entire response's credibility.

Verification principles and workflow for citations added during the rebuttal writing process.


Why Citations in Rebuttals Are High-Risk

Rebuttal citations are uniquely dangerous compared to paper citations:

  1. Short turnaround: Rebuttal period is 2–4 days. Authors write quickly under pressure.
  2. LLM suggestions are plausible-sounding but often wrong: AI-generated citation suggestions have an estimated 40% error rate — wrong author, wrong year, wrong title, or entirely fabricated.
  3. Reviewers check: An Area Chair or reviewer who discovers a fabricated citation in a rebuttal will immediately lose confidence in the entire response.
  4. No revision opportunity: Unlike a paper, a submitted rebuttal cannot be revised after submission.

Core rule: Never add a citation to a rebuttal response unless you have personally verified it exists via search.


When to Verify

Verify any citation that was:

  • Suggested by an LLM during Stage 2 drafting
  • Recalled from memory without looking it up
  • Copied from a reviewer comment (reviewers can also be wrong)
  • Added to counter a reviewer's "missing baseline" claim
  • Introduced to support a new claim not already cited in the paper

Existing citations from the paper itself generally do not need re-verification — but do confirm the claim you are citing actually appears in that work.


Verification Workflow

Step 1: Search

Use a reliable academic search engine:

Recommended search order:
1. Google Scholar (scholar.google.com)
2. Semantic Scholar (semanticscholar.org)
3. arXiv (arxiv.org) — for preprints
4. ACL Anthology (aclanthology.org) — for NLP/ACL papers
5. OpenReview (openreview.net) — for ICLR/ICML papers

Search query pattern:

"[Paper title keywords]" [first author last name] [approximate year]

Step 2: Confirm Existence

Verify the paper exists by confirming:

  • Title matches (allow minor capitalization differences)
  • At least the first author matches
  • Year is within ±1 of what you recall (preprints often differ from camera-ready year)
  • Venue matches what you intend to cite (workshop ≠ main conference)

Step 3: Verify the Specific Claim

When citing a specific finding, do not trust your memory of the paper's conclusion.

  • Locate the specific sentence or figure that supports your claim
  • If you cannot access the full text, cite more conservatively ("as suggested by [author]") or drop the citation
  • Do not paraphrase a result you have not read

Step 4: Record the Verified Reference

Format for use in rebuttal prose (no BibTeX needed in most rebuttal portals):

[Author et al., Year, "Title", Venue]
e.g.: [Chen et al., 2023, "LoRA: Low-Rank Adaptation of Large Language Models", ICLR]

When You Cannot Verify

If after searching you cannot confirm the paper exists:

Option A — Drop the citation If the claim can stand without it, remove the reference.

Option B — Hedge the claim "Prior work on [topic] (e.g., [general area]) suggests…" without citing a specific paper.

Option C — Mark for human verification If working with an LLM and you want to flag it:

[CITATION NEEDED — verify before submission: "[paper title you were trying to cite]"]

Never submit a rebuttal with a citation you have not verified, even if it sounds plausible.


Rebuttal-Specific Citation Patterns

Citing Your Own Prior Work

  • Verify the venue, year, and paper number match what appears in your submission's reference list
  • Consistency with the submitted paper matters — reviewers notice discrepancies

Citing a Reviewer-Suggested Baseline

  • Reviewers sometimes suggest comparing with papers that are not directly comparable
  • Verify the suggested paper actually evaluates on your task/dataset before agreeing to compare
  • You may respond: "We have examined [Paper], which evaluates on [different setting], making direct comparison difficult."

Citing to Defend Against "Missing Related Work"

  • Search first — the paper may genuinely exist and you missed it
  • If it exists, consider acknowledging it: "Thank you for pointing out [Paper]; we will add a discussion in Related Work."
  • If it does not exist (after searching), respond diplomatically: "We searched but could not locate this work; we would appreciate a link or full title."

Quick Verification Table

Citation SourceRisk LevelAction
LLM-suggested referenceVery highAlways verify before including
Recalled from memoryMediumVerify title and year
Copied from reviewer commentMediumVerify existence; reviewer may be wrong
Already in paper's reference listLowConfirm the claim is in that paper
Found via search in last 24hLowCross-check author/venue

Adapted from Claude Scholar's citation-verification skill for the rebuttal context. Original: https://github.com/Galaxy-Dawn/claude-scholar/blob/main/skills/citation-verification/SKILL.md

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Summarize extracted paper text into concise Markdown memory for later Stage 2 and Stage 4 background use, with fixed section headings and no fabricated claims.

日本語の概要は準備中です。原文の説明を表示しています。

runtsang/RebuttalStudio2182026年4月4日 更新

Systematic quality check for a completed rebuttal document before submission. Use after Stage 3 document compilation or after Stage 5 final remarks writing to verify completeness, tone, factual accuracy, and structural integrity. Catches common rebuttal errors before they reach reviewers.

日本語の概要は準備中です。原文の説明を表示しています。

runtsang/RebuttalStudio2182026年4月4日 更新

Systematic review response strategy guide for RebuttalStudio. Use when developing response strategies for reviewer comments, deciding how to classify concerns, or choosing between Accept/Defend/Clarify/Experiment approaches at any stage of the rebuttal pipeline.

日本語の概要は準備中です。原文の説明を表示しています。

runtsang/RebuttalStudio2182026年4月4日 更新

skills

無料

Multi-stage rebuttal analysis skill for RebuttalStudio. Use when organizing reviewer comments into stage-specific conference workflows, including stage1 breakdown, stage2 refinement, stage4 multi-round follow-up, and stage5 final remarks generation.

日本語の概要は準備中です。原文の説明を表示しています。

runtsang/RebuttalStudio2182026年4月4日 更新

Break down full ARR (ACL Rolling Review) reviewer responses into structured rebuttal units. Use when input contains ARR reviewer fields (Paper Summary, Strengths, Weaknesses, Comments/Suggestions) and numeric scores (Confidence, Soundness, Excitement, Overall Assessment, Reproducibility). Splits weaknesses and comments/suggestions into granular response items while preserving original wording for quoted issues.

日本語の概要は準備中です。原文の説明を表示しています。

runtsang/RebuttalStudio2182026年4月4日 更新

Base template for Stage1 reviewer-comment breakdown. Use when creating or extending conference-specific Stage1 skills (for example ICLR, ICML, NeurIPS, ACL) so shared extraction, splitting, and response-mapping logic stays consistent.

日本語の概要は準備中です。原文の説明を表示しています。

runtsang/RebuttalStudio2182026年4月4日 更新

runtsang のスキルをすべて見る

このスキルの問題を報告する