NeurIPS 2025 Review Released: Remember, a score is not your worth as a researcher!

[root]

"In every peer review cycle, some will win, some will lose, and most will need a group hug."
— Every CS author in late July

---

The Reviews Have Landed — Now What?

Today, the NeurIPS 2025 reviews have dropped for both the Main Track and Datasets & Benchmarks Track. Congratulations, commiserations, and caffeine to all! Whether you’re celebrating 6s or nursing some bruising 2s, the next week is your time to shine (or, at least, to clarify, correct, and convince).

Rebuttal period basics:

• Preliminary reviews are now available in your Author Console.
• You have until July 30th, 11:59pm AoE to submit your rebuttal.
• Reviewers see your responses starting July 31st.
• Discussion with reviewers: July 31st – August 6th, also AoE.

Golden Rules for Rebuttal:

• Be factual, focused, and respectful. Address factual errors, direct reviewer misunderstandings, and answer questions — don’t vent or include new experiments (save those for your next revision).
• Do not include: links to external pages, identifying information, or anything that violates the double-blind policy.
• If you get a low-quality, disrespectful, or inaccurate review: Use the “Author AC Confidential Comment” — but only for real issues, and proofread carefully.
• Public release: For accepted (and opt-in rejected) papers, your reviews and rebuttals will eventually go public. So, don’t say anything you wouldn’t want your PI, future employer, or your mom to read.

Pro tip: Don’t treat the average score as a crystal ball. ACs care more about review content and discussion than numbers alone.

---

Score Situation: What Are People Seeing?

This year, review scores feel lower than ever. Even the “inflated” public polls are showing a stark drop.

“Even with some upward bias in online voting, the proportion of papers with an average score below 3.5 is shockingly high. The days of 4 = borderline are over; now, 3 is the median.” — Xiaohongshu user

From the Xiaohongshu/WeChat stats (see image below in your post):

Score Range	Number of community votes
[4.33, 6]	117
[3.67, 4.33)	222
[3, 3.67)	324
[2, 3)	110

Interpretation:

• The most common range is [3, 3.67) — a “borderline” area that leaves authors anxious and ACs with tough choices.
• Less than 20% of authors are reporting means above 4.33.

Social media stories are awash with:

• Scores like 4/4/3/2 (“Should I rebut?”), 5/3/2/2, 4/4/4/2, or the heartbreaking all-2s.
• Anxiety about outlier reviewers (especially confident low scores).
• A common refrain: if your average is near 3, don't lose hope — but prepare a strong, factual rebuttal.

Reddit, Zhihu, and WeChat are full of batch statistics:

• “Chair total 14: 1 over 4.0, 4 between 3.0 and 4.0, 9 at or below 3.0.”
• “Scores 5/4/3/2 with one harsh but easily rebuttable review — do I have a shot?”
• Senior AC on X: “Out of my 100-paper batch: 1≥5.0, 6≥4.5, 11≥4.0, 25≥3.75, 42≥3.5. Naive cutoff for acceptance: ~3.75.”

In short:

• 3–4 is the new “limbo.”
• ≥4.5 is relatively rare, and usually a good sign.
• Outliers (both low and high) abound, and the rebuttal can make a difference if you’re close to the line.

---

Collected Score Table — 2025 Social Media Reports (up till when this article is written)

Here's a summary table of self-reported scores and cases seen on Chinese forums, Reddit, and WeChat:

Score Pattern	Track	Commentary
4/4/4/5	D&B/Main	Excellent, but still nerves about outliers
5/3/2/2	Main	Mixed; single strong positive, but two low (see if rebuttal helps)
4/4/3/2	Main	Very common; hope if the “2” can be addressed
3/3/3/3	Main/D&B	“Is this death?” — Actually, still possible with strong text
4/4/4/2	D&B	Confident negative reviewer; focus rebuttal there
4/3/3/3	Main	Borderline; ACs will look closely at review text
5/4/3/3/2	Main	5-reviewer paper; chance if low scores are weakly justified
2/2/2/2	Main	Unlikely, but never zero hope
4/5/2/2	Main	Outliers matter; address negatives, emphasize positives
4/3/2/5	Main	Again, a split batch — rebuttal is key
3/3/4/4	Main	With reviewer hints of raising score, good chance

Trends from these numbers:

• Most papers cluster between 3.0 and 4.0.
• It’s rare but not impossible to get four “strong accepts.”
• Negative reviews (scores 1–2) are often outliers, sometimes given with high confidence, but not always fatal if addressed with a clear, factual rebuttal.
• “Rebut or run?” is the most common question, and the answer is: if you are above 2.5, you should rebut unless all reviewers are negative with strong justifications. The bar for “safe accept” is much higher than before.

---

🧐 The Author Survey — And Why It Matters

NeurIPS 2025 has rolled out an optional survey asking authors to rank their own submissions. This isn’t a trap — it’s part of a longer-term research initiative to improve the peer review process, not to influence this year’s outcome.

How does it work?

• You’re asked to honestly rank your own submissions by their scientific contribution.
• The survey explores the relationship between author perceptions, review scores, and future impact.
• The ultimate aim? To find out whether author-informed rankings can improve the reviewing and decision-making process, leveraging insights from the isotonic mechanism.

Privacy note: Your response is completely anonymous, not shared with ACs or reviewers, and will be deleted after the study finishes.

Echoing the CSPaper Mission

This is very much aligned with the mission of CSPaper Review:

• Bringing transparency and feedback early in the process
• Experimenting with ways to make reviewing more fair, consistent, and author-centric
• Building a community that shares insights, pain points, and best practices

CSPaper Review already served 7,000+ unique users, processed 15,000+ reviews, and directly addresses the exact kind of reviewing bottleneck and author pain the community faces right now.

---

Why Try CSPaper Review? A New Perspective

If you’re feeling battered by NeurIPS reviews — or just want a second opinion for your next round (ICLR, AAAI, EMNLP...), CSPaper Review offers:

• Fast, conference-specific peer reviews (within 1 minute)
• Benchmarked realism (using real review data from OpenReview/social media)
• Feedback on where your paper might stand in a different venue
• Useful for diagnosing common reviewer objections, clarifying writing, or testing rewrites before next submission
• Free to use (up to 20 reviews/day), no login required

Remember, tools like CSPaper Review won’t change your current NeurIPS fate, but they can help diagnose weaknesses, frame your rebuttal, and plan for future resubmission or revision.

CSPaper Team is also keen to learn how our tool matches the reviews you obtain, we welcome you to share with us or the community by emailing us support@cspaper.org or create a post here!

---

End Words: Keep Calm, Rebut Wisely, and Stay in the Game

To everyone feeling crushed, furious, or numb: You are not alone.

• Most people’s scores are “mid.”
• Outliers are everywhere.
• The review text matters more than the mean.
• “Borderline” is the new normal.

Do your best rebuttal — factual, concise, and polite.
Discuss with your co-authors, check the CSPaper Review tool for a fresh look, and don’t be afraid to ask your community for support or advice.
And remember: a score is not your worth as a researcher.

May the review gods be with you. May your scores rise, and may you touch grass before the notifications drop.

— The CSPaper.org team

---

Now, go forth and rebut. Or, at least, go outside for five minutes — you earned it.

[cqsyf]

As a reviewer, I got messages from OpenReview about authors withdraw their submission upon seeing the reivew scores. The two withdrawn papers have scores as follows:

4,3,3,3
4,3,3,2

I guess, mean score >=4 would be a good spot for a chance of acceptance? anything below could be kinda far from acceptance?

[Joanne]

Who’s Adam? The Most Ridiculous NeurIPS Review Is Here

Have you received your NeurIPS 2025 review results yet? Brace yourselves—it’s officially open season for venting frustrations about reviewer comments!

Right on cue, we stumbled upon what might just be the funniest and most outrageous NeurIPS comment of the year. Credit goes to Yiping Lu, an assistant professor in Northwestern University’s Department of Industrial Engineering and Management Sciences and a proud Peking University alum.

Lu posted a screenshot of this legendary review on X (formerly Twitter), and it immediately went viral—over 505,000 views in less than a day! Make your guess is this review from LLM?

The Reviewer’s Baffling Comment

Line 336: “Both architectures are optimized with Adam.”
“Who/what is ‘Adam’? I think this is a very serious typo that the author should have removed from the submission.”

Yes, folks, you heard it correctly. The reviewer, presumably an expert, genuinely believed “Adam”—one of the most fundamental optimization algorithms in machine learning—was either a typo or possibly a mysterious colleague named Adam.

Community Roast

Even Dan Roy, a respected professor and machine-learning researcher, couldn’t resist firing a shot:

“NeurIPS reviews are complete trash.”

🧨 The Deeper Issue: Quantity vs. Quality

With NeurIPS submissions skyrocketing toward 30,000 papers this year, it’s becoming glaringly obvious that human reviewers alone can’t handle the load. This mismatch between submission volumes and reviewer capacity inevitably leads to bizarre review outcomes like the infamous “Adam incident.”

Could AI Save Peer Review?

AI tools are already creeping into academic peer review. UC Berkeley postdoc Xuandong Zhao recently tweeted:

“#NeurIPS2025 reviews are out, and the authenticity of reviews surprises me again.

Two years ago, maybe 1/10 felt AI-assisted. Now? It seems 9/10 are AI-modified, going beyond simple grammar fixes to fully generated comments.

As someone researching AI-generated content detection, these might be impressions rather than hard data. Still, this trend worries me: AI writing papers, AI reviewing them, AI publishing them. Is this really the future of academia we want?”

From drafting manuscripts to reviewing and publishing, AI now permeates the entire academic pipeline.

What Now? It’s Rebuttal Season!

As funny as this review is, it’s still critical to tackle it seriously in your rebuttals. Thankfully, one helpful community member shared an invaluable resource from 2020:

How We Write Rebuttals

Final Note

Don’t forget to share your story by contributing to our cspaper.org!

[Sylvia]

NeurIPS Author-Reviewer Discussion: Key Points

• Purpose:
  Only clarify reviewer questions. Do not add new arguments or extensive commentary.

• How to Respond:

  • Use the “Official Comment” button for each discussion thread.
  • Be brief and to the point.
  • Do not ask or urge reviewers to reply; ACs/PCs will manage that.

• Confidential Comments:

  • Use “Author AC Confidential Comment” for private notes to ACs (not visible to reviewers).
  • Available until Aug 6.

• Anonymity & Conduct:

  • Do not reveal your identity or use links.
  • Remain respectful at all times.

• Timeline:

  • Discussion closes Aug 6, 11:59pm AoE.
  • Reviewer final ratings/justifications are hidden until notification.
  • Save reviews before Aug 6 if you plan to withdraw.

• Navigation:
  Use filters in OpenReview to view messages by author/reviewer/AC.

[root]

Author-Reviewer discussion is extended!

• Author-Reviewer discussion is extended by 48 hours, now closing on Aug 8, 11:59 PM AoE.
• AC-Reviewer discussions remain unchanged and will conclude by Aug 13.
• If a reviewer hasn’t replied to your rebuttal, feel free to initiate discussion and tag your AC for support.
• Even if a reviewer has marked “Mandatory Acknowledgement,” you’re encouraged to continue engaging if open questions remain.
• Late rebuttals (posted entirely after Jul 30) will be ignored to ensure fairness.
• You won’t see updated ratings until decisions are released. Reviewers can still express if your rebuttal was helpful.

Good luck and happy discussing!

[RiskLimp]

Hey folks,

This is a quick but important reminder from the NeurIPS 2025 official email for everyone reviewing

It lists some patterns emerge that could lead to red flags and future consequences for reviewers. Let's break it down.

---

TL;DR – Don’t Click and Ghost

Clicking “Mandatory Acknowledgement” does NOT excuse you from discussion participation.

• Reviewers must actively participate in author discussions before clicking the acknowledgement button.
• Those who don’t engage will be flagged by Area Chairs (ACs) and the system.
• Such flags could lead to penalties under NeurIPS' Responsible Reviewing policy and may impact future invitations.
• Late rebuttals (posted in full after July 30) are to be ignored.
• Author-reviewer discussions have been extended to Aug 8, 11:59pm AoE. AC-reviewer discussions end Aug 13.

---

Red Flags to Avoid

These are specific behaviors being monitored; and strongly discouraged:

• Clicking "Mandatory Acknowledgement" but not posting anything in the discussion thread
• Ignoring rebuttals or failing to respond to author comments
• Waiting until the very last moment to engage (or not engaging at all)
• Giving vague positive feedback in discussion but a harsh final justification
• Silent rating changes without explanation or communication
• Assuming low-scored papers don’t deserve replies (they do!)

---

What You Should Do Now

• Log in to OpenReview
• Read all rebuttals and comments
• Post a reply, even if your position hasn't changed:
  • If concerns are resolved: say so.
  • If not resolved: explain why.
• Ask clarifying questions if needed.
• Keep your final justification consistent with your public discussion.
• If you're unsure whether you need to respond, ask your AC.

---

When Is It OK to Submit “Mandatory Acknowledgement”?

Only when you have:

1. Read the author rebuttal
2. Engaged in the author-reviewer discussion
3. Completed your final justification and updated your rating (if needed)

---

️ Responsible Reviewing Is Active

NeurIPS 2025 is running its Responsible Reviewing initiative:

• ACs can flag reviewers for lack of participation
• Negligent reviewing may impact your own submissions
• A record of poor reviewing will be maintained for future years

---

🧠 Final Thought

Let's treat authors with:

• Fairness
• Politeness
• Calmness
• Attention
• Focus on scientific merit

Whether you're supporting a paper or recommending rejection, a few sentences go a long way in building a healthy review culture.

[root]

This Year’s Community Observations

• Reviewer Silence: Most authors receive little to no real engagement — sometimes only an MA.
• Score Disappearance: If a reviewer updates their rating/final justification, their score vanishes from the author console until decisions.
• Community Frustration: Many feel the “discussion period” is a one-way street; reviewers rarely debate or clarify further.
• Randomness: Even with identical scores, some papers get in, others don’t. “AC lottery” and “reviewer lottery” are real!
• Field Differences: Some subfields have lower average scores; a 4.0 may be top-20% in a tough area, borderline in others.

---

Exhaustive Table: Score Scenarios & Acceptance Hope

Below is a table constructed from real examples in the community (Zhihu, Reddit, etc.).

Case	Initial Scores (C=Confidence)	Actions During Rebuttal	What Happened After	Community Analysis	Acceptance Hope
A	5,5,4,3	One 4 promised to increase; 3 is silent	Avg. visible 4.67 or 4.0 (if one vanishes)	4+ avg with one “up” and no strong negative	Very good chance
B	5,4,4,2	2 was highest confidence; author tried to convince; 4 promised up	Score vanishes after “up”, avg. jumps if visible	If both “up,” strong chance. If 2 stays, borderline; depends on AC
C	4,4,4,4	All 4s; only MA or silence	Avg 4.0	No reject, but no clear “champion.” Depends on AC and field bar	50–75% (“bubble”)
D	5,5,3,2 (C: 3,3,5,3)	Tried to “fix” the 2; 3 was picky	Only MA; avg 3.75/4.0	Negative champion can kill it if not convinced; AC matters	Risky but possible if AC sympathetic
E	3,3,4,5	4 promised “up”; 3 did MA	Avg jumps (3.75–4.0)	If “up” confirmed, likely to be accepted, esp. if 3 is no longer counted	Good odds if “up” happens
F	3,4,4,4,5	4 said updated; 5 silent	Avg 4.2	Most positive, one “update” is good	Likely accept if no major issues
G	3,3,3,4	Only one 4, rest 3s, some silence	Avg 3.25	Borderline or below, need at least one “up”	Unlikely unless field has very low average
H	4,4,3,3	4s MA, 3s silent	Avg 3.5–4.0	Borderline, need at least one 3 to up	True “bubble”—AC can go either way
I	5,4,3,3 (C: 2,1,4,4)	Both 3s high confidence; all silent	Avg 3.75	Weak, but not hopeless. AC might tip if rebuttal strong.	30–50%
J	5,5,5,4,3	4 says will up, 3 silent	Avg >4	Excellent if 4 moves up, still good if not
K	4,4,3,2	All silent	Avg 3.25	Very hard; needs miracle up-vote from 2 or 3	Unlikely
L	5,4,3,2; up to 5,5,3,3	After rebuttal all up by 1	Avg now 4	If real, strong, but only if up-votes stick
M	3,5,5,2	Tried hard to move 2, 5 says up	If avg now 4+, possible	Must move 2 or AC must ignore	40%
N	4,4,4,3	All MA, no comments	Avg 3.75	Bubble, depends on AC or field
O	2,3,5,5	3/2 not responding, 5/5 happy	Avg 3.75	If no “up,” borderline; depends on AC and field	30–60%
P	3,4,5,5	Some up, some down	Avg 4.25	Good if up, depends if 3/4 “negative champion”
Q	2,2,2,2	Submitted rebuttal anyway	Avg 2	“No hope” zone
R	4,3,5,5	3 not convinced, 5/5 positive	Avg 4.25	Good odds, but 3 can kill if AC is negative
S	3,3,4,4 (C: 4,4,3,3), all said “up”	All promise up, scores vanish	Should be 4/4/4/4+	Very strong, likely to be accepted
T	4,4,4,5	MA only, no real engagement	Avg 4.25	No negative champion, likely safe
U	5,4,4,2; 2 up to 3, 4 to 5	Scores vanish, avg >4	Strong, very likely accept
V	5,4,4,3; MA from 4s, 3 silent	Avg 4.0	Borderline, need at least one “up”
W	5,4,3,2; 2 still visible after MA	2 unchanged	Low chance unless AC overrules
X	3,4,4,4	3 says “thanks, still weak”	Avg 3.75	Need 3 to up, else risky
Y	3,3,3,4,4	All 3s, one 4	Avg 3.4	Very difficult
Z	5,5,4,2	2 is negative champion	Avg 4	On the line; needs AC intervention

---

What the Community Says

• Silence ≠ Rejection: Most reviewers do not reply, especially to borderline or mid-range scores. Don’t panic!
• MA Means Finalized: Once you see “Mandatory Acknowledgement,” you can’t see the score if they updated it, but they may have changed it up or down.
• Upvotes = Hope: Any reviewer explicitly saying “I am raising my score” is good; if the score vanishes, assume they did.
• Downvotes Happen: Occasionally, a reviewer will lower after rebuttal; this is rare but possible if the rebuttal is weak or reveals flaws.
• “Negative Champion” Danger: A single reviewer (usually the lowest) with high confidence can block a paper if not convinced. Always try to address their points specifically.

Area Chairs (AC) Role

• ACs often have leeway in borderline cases, especially if rebuttal is strong or all concerns are addressed except one.
• If all reviewers are “neutral” (3s and 4s), and no strong negative, ACs may push for acceptance, especially if field is tough.
• Community advice: if you have real engagement from at least one reviewer, and no strong rejections, odds are decent.

---

Finally

Here are the main takeaways from recent statistics (see screenshots):

• Acceptance Rate: NeurIPS acceptance rate has hovered around 25% for the last five years (see Paper Copilot NeurIPS 2025 statistics).
• Score Distribution: Most “active” (not withdrawn) papers have initial means in the 3.0–4.2 range.
• Community feedback is consistent with these tools: avg. 4+ is usually "safe", 3.5–4.0 is a coin toss, <3.5 is very unlikely.
• Confidences help: 4s and 5s from high-confidence reviewers are