Skip to content

Reviewer Perspective Guide

Table of Contents

Four-Dimensional Review Criteria

CriterionWhat Reviewers Look ForCommon Deduction Reasons
QualityTechnical soundness, well-supported claimsFlawed proofs, missing baselines, unfair comparisons
ClarityClear writing, reproducible by expertsAmbiguous notation, missing details, poor organization
SignificanceCommunity impact, advances understandingIncremental improvement, narrow applicability
OriginalityNew insights (doesn't require new method)Known results, obvious extensions

Six-Point Scoring Scale

(NeurIPS scale, widely adopted)

ScoreRatingTypical Characteristics
6Strong AcceptGroundbreaking contribution, technically flawless, very high impact
5AcceptTechnically solid, high impact, well-written
4Borderline AcceptSolid contribution, limited evaluation or novelty concerns
3Borderline RejectTechnical merits but weaknesses outweigh strengths
2RejectTechnical flaws, insufficient evaluation, unclear contribution
1Strong RejectKnown results, ethical issues, or fundamental errors

What Moves Papers from 3→4 (Borderline → Accept):

  • [ ] Address obvious weakness proactively (limitations section)
  • [ ] Add one more strong baseline or ablation
  • [ ] Improve clarity of main contribution statement
  • [ ] Add reproducibility details (code, hyperparameters)

Reviewer Reading Behavior

Understanding how reviewers read helps prioritize your effort:

Paper Section% Reviewers Who ReadTime SpentImplication
Abstract100%2-3 minMust be perfect
Introduction90%+ (skimmed)3-5 minFront-load contribution
Figures/TablesBefore methods2-3 minFigure 1 is critical
MethodsOnly if interested5-10 minDon't bury the lede
ExperimentsIf methods seem sound5-10 minClear claims + evidence
AppendixRarely (<30%)As neededSupplementary only

Key Insight: If your abstract and intro don't hook reviewers, they may never read your methods.

Pre-Submission Checklist

Universal Checklist (All Venues)

  • [ ] Paper compiles without errors
  • [ ] All figures are referenced in text
  • [ ] All tables are referenced in text
  • [ ] No orphaned citations (every \cite has a bib entry)
  • [ ] No placeholder text (TODO, FIXME, XXX)
  • [ ] Anonymous submission (no author names in blind review)
  • [ ] Page limit respected (excluding references)
  • [ ] Consistent notation throughout
  • [ ] All acronyms defined on first use
  • [ ] Limitations section included

NeurIPS Specific (16 items)

  • [ ] Paper checklist completed (Appendix)
  • [ ] Broader Impact Statement included
  • [ ] Code submission prepared (if applicable)
  • [ ] Lay summary prepared (for accepted papers)
  • [ ] Ethics review flagged if applicable
  • [ ] Supplementary material ≤ 50MB
  • [ ] Main paper ≤ 9 pages (+ unlimited references/appendix)
  • [ ] Uses official NeurIPS style file
  • [ ] Reproducibility details: random seeds, compute, datasets
  • [ ] Error bars included with methodology specified
  • [ ] Statistical significance tests where appropriate
  • [ ] Dataset licensing and consent documented
  • [ ] Potential negative societal impacts discussed
  • [ ] Limitations clearly stated
  • [ ] Comparison with appropriate baselines
  • [ ] Ablation studies for key design choices

Rebuttal Best Practices

Dennett's Criticism Method (Adapted for Academic Rebuttals)

  1. Acknowledge: Re-state the reviewer's concern fairly
  2. Agree: Find points of agreement
  3. Learn: Show what you learned from the criticism
  4. Respond: Then (and only then) address disagreements with evidence

Rebuttal Structure Template

We thank Reviewer [X] for their thoughtful feedback.

**[Concern 1: Brief summary]**

We agree that [acknowledgment]. To address this:
- [Action taken / experiment added / clarification provided]
- [Evidence: "Table X shows..." / "We added Section Y..."]

**[Concern 2: Brief summary]**

[Same structure...]

**Summary of Changes:**
- [Bullet list of all modifications]

Common Rebuttal Mistakes

  • ❌ Arguing that the reviewer is wrong
  • ❌ Ignoring a concern (address ALL points)
  • ❌ Promising future work instead of showing results
  • ❌ Being defensive or dismissive
  • ✅ Providing concrete evidence (new experiments, tables)
  • ✅ Acknowledging valid criticisms
  • ✅ Showing actual revisions made

Released under the MIT License.