Review Criteria

Review Criteria

Unified review criteria for paper audit, based on major conference standards (NeurIPS, ICLR, ICML, IEEE, ACM).

Four-Dimensional Review Framework

1. Quality (Weight: 30%)

What reviewers assess:

• Technical soundness of claims and proofs
• Correctness of mathematical derivations
• Fairness of experimental comparisons
• Adequate baselines and ablation studies
• Reproducibility of results

Common deduction reasons:

• Flawed proofs or incorrect derivations
• Missing baselines or unfair comparisons
• Insufficient experimental evaluation
• Claims not supported by evidence
• Missing error bars or statistical tests

Automated checks mapping: Bibliography verification, Logic & coherence analysis

2. Clarity (Weight: 30%)

What reviewers assess:

• Clear and precise writing
• Logical document organization
• Consistent notation throughout
• Reproducible by domain experts
• Appropriate use of figures and tables

Common deduction reasons:

• Ambiguous notation or undefined terms
• Poor organization or missing sections
• Long, convoluted sentences
• Missing method details
• Figures without proper captions or references
• Verbose or AI-generated caption text (e.g., "The figure shows") or inconsistent casing

Automated checks mapping: Format check, Grammar analysis, Sentence complexity, De-AI detection, Figure/table references, Consistency

3. Significance (Weight: 20%)

What reviewers assess:

• Impact on the research community
• Advances understanding of the problem
• Practical applicability
• Non-incremental contribution

Common deduction reasons:

• Incremental improvement over existing work
• Narrow applicability
• Limited novelty
• No clear advantage over simpler methods

Automated checks mapping: Logic & coherence analysis (methodology justification)

4. Originality (Weight: 20%)

What reviewers assess:

• New insights or perspectives
• Novel methodology or approach
• Creative problem formulation
• Not obvious extension of prior work

Common deduction reasons:

• Known results or obvious extensions
• Insufficient differentiation from prior work
• Missing discussion of novelty
• AI-generated content without original thought

Automated checks mapping: De-AI detection

Scoring Algorithm

For each dimension D:
    base_score = 6.0
    for each issue mapped to D:
        if severity == "Critical": score -= 1.5
        if severity == "Major":    score -= 0.75
        if severity == "Minor":    score -= 0.25
    dimension_score = max(1.0, score)

overall = quality * 0.30 + clarity * 0.30 + significance * 0.20 + originality * 0.20

Score-to-Label Mapping

Important Notes

• Automated scores cover Clarity well but have limited ability to assess Quality, Significance, and Originality
• LLM judgment supplements automated checks for the latter dimensions
• Scores should be treated as indicators, not definitive assessments
• Always prioritize Critical issues regardless of overall score