AgentSkillsCN

review-paper

模拟一次严谨的 ACL 式同行评审，对论文进行多轮平行分析，并综合形成评审报告。在提交前使用此功能。

SKILL.md
--- frontmatter
name: review-paper
description: Simulate a thorough ACL-style peer review of the paper. Runs parallel analysis passes and synthesizes a reviewer report. Use before submission.
context: fork
agent: task

ACL Paper Review Simulation

Perform a comprehensive review of the LaTeX paper in the paper/ directory, simulating an ACL program committee reviewer.

Phase 1: Read the Full Paper

Read every .tex file in order. Understand the complete argument before critiquing.

Phase 2: Parallel Analysis Passes

Run these checks (as subagents where possible):

Pass A — Novelty & Positioning

  • •What is the core contribution? Is it clearly stated?
  • •How does it differ from prior work? Is the related work section thorough?
  • •Are there obvious missing baselines or comparisons?
  • •Would an ACL reviewer find this sufficiently novel?

Pass B — Technical Soundness

  • •Are the claims supported by evidence?
  • •Is the experimental setup described in enough detail to reproduce?
  • •Are the evaluation metrics appropriate for the task?
  • •Are statistical significance tests used where appropriate? (CIs, p-values, bootstrap)
  • •Do the ablations isolate the claimed contributions?
  • •Are there confounds the authors haven't addressed?

Pass C — Clarity & Presentation

  • •Is the abstract self-contained and informative?
  • •Does the introduction motivate the problem effectively?
  • •Is the paper well-organized? Does the argument flow logically?
  • •Are figures and tables informative and well-designed?
  • •Is notation consistent and clearly defined?
  • •Are there sections that are too dense or too sparse?

Pass D — Numerical Consistency

  • •Do numbers in the text match their source tables?
  • •Are percentages computed correctly from the underlying counts?
  • •Do reported improvements (e.g., "+3.2%") match the actual difference between numbers?
  • •Are baselines compared fairly (same data splits, preprocessing)?

Pass E — ACL Compliance

  • •Within page limits (8 pages + unlimited references for long, 4 for short)?
  • •Anonymous (no author names, no self-identifying URLs, no "our previous work [1]")?
  • •Required sections present (Abstract, Introduction, Related Work, Methodology, Experiments, Conclusion)?
  • •Ethics statement if needed?
  • •Limitations section (required for ACL 2024+)?

Phase 3: Synthesize Review

Format as an official ACL review:

code
OVERALL SCORE: [1-5] (1=reject, 3=borderline, 5=strong accept)
CONFIDENCE: [1-5]

SUMMARY:
[2-3 sentence summary of the paper and its contribution]

STRENGTHS:
S1. [specific strength with reference to section/table]
S2. ...
S3. ...

WEAKNESSES:
W1. [specific weakness with suggested fix]
W2. ...
W3. ...

QUESTIONS FOR AUTHORS:
Q1. ...
Q2. ...

MISSING REFERENCES:
- [Any related work the authors should cite]

MINOR ISSUES:
- [Typos, formatting, small suggestions]

RECOMMENDATION:
[Specific actionable feedback: what would move this from a 3 to a 4?]