CSPaper: sidekick of peer review

Reproducibility is key in science, but in computer science and machine learning (ML), it's often overlooked. Even though most ML experiments should be able to reproduce comparatively easy (except for some papers that require very heavy computation, many top papers can't be reproduced. Often, the code is not even has very basic and obvious errors in it, being far away from even being re-used. Here a more elaborate blog post on this: https://www.mariushobbhahn.com/2020-03-22-case_for_rep_ML/ This difference raises questions: Should CS and ML have stricter rules for reproducibility in peer reviews? Should more CS and ML researchers retract papers that can't be reproduced? How can we encourage researchers to make their work reproducible? One effort to fix this is "Papers Without Code," where people can report ML papers that can't be reproduced. Its creator says, "Unreproducible work wastes time, and authors should make sure their work can be replicated." Improving reproducibility could greatly help peer reviews. If reviewers could easily test the results in a paper, they could: Check if the results are correct Understand the work better Give more helpful feedback Spot potential problems early This would lead to better quality research being published. It would also save time and resources in the long run, as other researchers wouldn't waste effort trying to build on work that doesn't actually work. What do you think? How can we make CS and ML research more reproducible? How would this change peer reviews?

The Verdict: ACL 2025 Review Scores Decoded This year’s Overall Assessment (OA) descriptions reveal a brutal hierarchy: 5.0 "Award-worthy (top 2.5%)" ️ 4.0 "ACL-worthy" 3.5 "Borderline Conference" 3.0 "Findings-tier" (Translation: "We’ll take it… but hide it in the appendix") 1.0 "Do not resubmit" (a.k.a. "Burn this and start over") Pro tip: A 3.5+ OA avg likely means main conference; 3.0+ scraps into Findings. Meta-reviewers now hold life-or-death power—one 4.0 can save a 3.0 from oblivion. Nightmare Fuel: The 6-Reviewer Special "Some papers got 6 reviewers—likely because emergency reviewers were drafted last-minute. Imagine rebutting 6 conflicting opinions… while praying the meta-reviewer actually reads your response." Rebuttal strategy: 2.0? "Give up." (Odds of salvation: ~0%) 2.5? "Worth a shot." 3.0? "Fight like hell." The ARR Meat Grinder Just Got Worse New changes to the ARR (Academic Rebuttal Rumble): 5 cycles/year now (April’s cycle vanished; June moved to May). EMNLP’s deadline looms closer — less time to pivot after ACL rejections. LLM stampede: *"8,000+ submissions per ARR cycle! "Back in the days, ACL had 3,000 submissions. No Findings, no ARR, no LLM hype-train. Now it’s just a content farm with peer review." How to Survive the Madness Got a 3.0? Pray your meta-reviewer is merciful. 🤬 Toxic review? File an "issue" (but expect crickets). ARR loophole: "Score low in Feb? Resubmit to May ARR and aim for EMNLP." The Big Picture: NLP’s Broken Incentives Reviewer fatigue: Emergency reviewers = rushed, clueless feedback. LLM monoculture: 90% of papers are "We scaled it bigger" or "Here’s a new benchmark for our 0.2% SOTA." Findings graveyard: Where "technically sound but unsexy" papers go to die. Final thought: "If you’re not gaming the system, the system is gaming you." Adapted from JOJO极智算法 (2025-03-28) Share your ACL 2025 horror stories below! Did you rebut or run?

Hi everyone, I recently read a thought-provoking paper from EMNLP 2023 titled "Automatic Analysis of Substantiation in Scientific Peer Reviews" by Guo et al., and I think it’s worth bringing into our community for discussion. The Problem: Peer Review Quality is Declining [image: Untitled-design-2-1.png] If you’ve submitted to AI/ML conferences lately, you might’ve experienced reviews that feel vague, generic, or just unhelpful. You’re not alone. The paper highlights a concerning trend: the substantiation level in reviews, how well claims are supported by evidence, has been declining in major NLP conferences over the past few years. This is likely due to the exploding number of submissions and a shortage of expert reviewers. Combine that with tight deadlines and unclear review guidelines, and you get a perfect storm for poor reviewing practices. “The proportion of supported claims in reviews dropped steadily from CoNLL 2016 to ARR 2022.” — Guo et al. 2023, EMNLP Findings ️ The Proposed Solution: argument mining for review analysis To address this, the authors developed a novel argument mining system that automatically extracts claim-evidence pairs from peer reviews. They even created a dataset called SubstanReview with 550 annotated reviews and introduced a metric called SubstanScore — a quantifiable way to measure review quality based on substantiation. Highlights: Defines a new NLP task: claim-evidence pair extraction in peer reviews Introduces SubstanScore: % of claims backed by evidence × review length Benchmarks fine-tuned transformers like RoBERTa and SpanBERT Shows ChatGPT underperforms on this task even with detailed prompts This is not just a cool NLP task; it has real implications for how our scientific community maintains quality and trust. Questions for the Community Should conference chairs integrate automated quality checks (like substantiation analysis) into review processes? Would you support a "review scorecard" that flags unsubstantiated or low-quality reviews? How can we balance automation with fairness, given that substantiation doesn’t capture other dimensions like factuality or expertise? Is it time we start training reviewers explicitly, using tools like SubstanReview as part of reviewer onboarding? 🧭 Why This Matters AI/ML conferences are the heartbeat of our field, and if the gatekeeping mechanism (i.e., peer review) starts to wobble, everything else follows. This paper brings both diagnosis and prescription, and while it’s not a silver bullet, it may just be the kind of infrastructure we need to restore faith in the system. Paper link: Automatic Analysis of Substantiation in Scientific Peer Reviews (EMNLP 2023) Dataset & Code: SubstanReview GitHub Repo

Just noticed that ICML 2025 has taken a small but meaningful step toward OpenReview: the reviews of accepted papers will eventually be made public. While this isn't full-fledged open review yet, it's a clear signal that change is coming. As a reviewer myself, I felt overwhelmed by the sheer volume of submissions this year. Unfortunately, I felt a noticeable drop in quality. Some papers were clearly submitted in a "let's try our luck" fashion. In this context, I sincerely hope that top AI/ML conferences will eventually follow ICLR's model and adopt fully open peer review. Why Open Review Matters For Reviewers: Knowing that reviews will be public adds a layer of accountability. It encourages more thoughtful, constructive, and responsible feedback. No more careless 1-scores or copy-pasted comments. For Authors: When reviews are public, authors will think twice before submitting undercooked ideas. Fear of negative reviews being visible online can act as a natural filter to avoid "lottery-style" submissions. For the Community: Public reviews help newcomers learn how to write better papers and better reviews. It also reduces the burden on reviewers caused by the "Fibonacci submission strategy" (endless revise-and-resubmits across top venues), and ultimately improves the quality of accepted papers. Final Thoughts Open review isn't a silver bullet, but in this era of exploding submission numbers, it’s a change worth pursuing. I hope to see more top-tier conferences move toward transparent and accountable reviewing, bringing the focus back to research quality, not just acceptance rates.

We’re excited to introduce a new category on cspaper.org: Anonymous Sharing & Supplementary Materials Purpose This category is designed specifically for anonymized sharing of supplementary materials such as: Additional experimental results or figures Extended ablation studies Links to datasets or demos Supplementary explanations that didn’t fit in the main paper It’s especially useful during rebuttals, when you may need to share extra content with reviewers but remain compliant with anonymity and strict page limits. How It Works Only unverified (but registered) users can post in this category: make sure to skip filling your email during registration. You can edit or delete your post anytime Use a username that doesn’t reveal your identity Share the link to your anonymous supplementary materials post/topic This keeps everything in line with double-blind peer review policies while giving you a legitimate way to share supporting materials. ️ Stay Anonymous Please do not use real names or affiliations when posting. If you're unsure how to create an anonymous username, try one of these generators: Jimpix Random Username Generator UsernameGenerator.com NordPass Username Generator Dashlane Username Generator Example Use Case You're writing the rebuttal for your paper submitted and reviewed by NeurIPS/ICML/ACL. You need to trim down your rebuttal to fit the rebuttal length limit. But, you want to share detailed results or code (causing violation of rebuttal length) with reviewers during the rebuttal. Just post them anonymously in that category and include the link in your response! Pro Tip you can make your url even shorter by leveraging services like TinyURL and Bitly. Start Posting Visit the category here and share your materials: https://cspaper.org/category/10/anonymous-sharing-supplementary-materials

Nature’s latest article dives into a topic that’s both timely and controversial for the Computer Science research community: How Artificial Intelligence is Changing Peer Review [image: d41586-025-00894-7_50775772.jpg] Why CS Researchers Should Care The peer review process is foundational to how we validate and publish research in venues like NeurIPS, ICML, CVPR, and ACL. Now, AI tools — including LLMs — are being used to: Detect issues in code, datasets, and statistics Polish academic writing and reviewer comments Generate structured reviewer feedback As developers and users of these technologies, we’re not just spectators — we’re stakeholders in how AI shapes peer review. 🧠 Highlights from the Article Automation meets evaluation: Editors are using AI to support reviewers, flag ethical issues, and accelerate turnaround times. Rising concerns: Dilution of human judgment and domain expertise Risk of leaking confidential submissions to third-party AI tools Over-reliance on black-box systems in scholarly assessment Calls for transparency: Journals are starting to draft policies, but the CS community still operates largely in a gray area. Open Questions for CS Conferences As AI becomes embedded in our research workflows, what happens to the review process itself? Should CS conferences allow or restrict AI-generated reviews? How do we audit the use of AI tools in paper writing and reviewing? Is it time to develop community guidelines for LLM usage in peer review? Join the Discussion Have you seen AI-assisted reviews in action? Used ChatGPT to clarify your own review comments? Let’s talk: 🧭 What role should AI play in the future of peer review? ️ How do we protect the integrity of double-blind reviewing? ️ Can we strike a balance between efficiency and ethics? Share your experiences and thoughts below. Source: Nature — How Artificial Intelligence is Changing Peer Review, March 27, 2025

Originally posted on Bluesky by Serge Belongie (@serge.belongie.com), this timely thread opens a conversation about decentralizing paper presentations at NeurIPS. 🧭 The Poll That Sparked a Discussion Serge Belongie (University of Copenhagen) and Søren Hauberg (DTU) are exploring a simple but profound question: Would you present your next NeurIPS paper in Europe instead of traveling to San Diego (US), if this was an option? They’re collecting data to gauge interest in offering official NeurIPS paper presentations at a European location, without requiring long-haul travel to the US. Link to participate in the poll (Google Form) 🧠 Why This Matters Several compelling motivations were outlined: ️ Environmental Impact: Transatlantic flights carry a heavy carbon footprint. Family Considerations: Long travels can be difficult for researchers with children. US Entry Concerns: Visa policies and border control have become increasingly strict and unpredictable. Conference Scale: NeurIPS has grown large — some argue, too large for meaningful in-person interaction. 🧩 Community Reactions The thread generated a flurry of responses from researchers around the world: Concerns about fragmentation: “Splitting the community weakens the social connections.” — Kosta Derpanis Necessity due to constraints: “Many US-based researchers fear not being let back in if they travel.” — David Marx Existing precedents: “Since the pandemic, we’ve had NeurIPS@Paris and CVPR@Paris satellite meetups.” — David Picard Institutional support growing: IJCAI 2025 will officially include a satellite event in Guangzhou, China to accommodate researchers unable to enter Canada. → IJCAI Announcement Feasibility discussions already underway: “There is discussion of a NeurIPS 2025 satellite outside the US but in a similar timezone.” — Thomas Dietterich Implications for Peer Review & Conference Design This idea challenges the default notion of “main venue” = “main stage.” As the global AI/ML community grows, inclusivity and accessibility in paper presentation become just as critical as peer-review rigor. Should presenting a NeurIPS paper require thousands of dollars and international visas? While decentralization may create new logistical and identity challenges, it could also lead to: More inclusive participation Better environmental stewardship Stronger regional AI communities What Do You Think? Would a European satellite venue dilute the prestige of NeurIPS, or increase its relevance and reach? Is it time for flagship CS conferences to rethink hybrid and distributed formats, especially post-pandemic? Drop your thoughts below. Special thanks to @serge.belongie.com for initiating this much-needed discussion.

The recent ICML 2025 review cycle has sparked outrage and dark humor across the ML community. Here’s a compilation of jaw-dropping anecdotes from Zhihu (China’s Quora) exposing the chaos, ranging from clueless reviewers to systemic failures. Buckle up! 1. The "k-NN" Incident User "写条知乎混日子" dropped this gem: "The reviewer asked me: ‘What does the ‘k’ in k-NN stand for?’" Yes, a reviewer at ICML, a top-tier ML conference, needed clarification on one of the most basic ML concepts. 2. The "Pro vs. RPO" Mix-Up User "CpGD7" shared: "The reviewer misread ‘rpo’ as ‘pro’ and questioned why our ‘advanced version’ lost to baselines. Next time, should I rename my main experiments ‘Promax’ to get accepted?" 3. The "I Didn’t Have Time to Check Proofs" Confession User "虚无", a reviewer, admitted: "I got assigned 5 theoretical papers. Checking proofs properly takes 7–10 days per paper. I only had time to verify the first two; the rest got high scores based on ‘intuition’ because I couldn’t validate the math." This raises a serious ethical concern: Papers are being accepted/rejected based on guesses, not rigor. 4. The "Citation Mafia" Reviewer User "Jane" reported: "A 1-star reviewer demanded we cite 7 unrelated papers — 6 of which were by the same author. We withdrew the submission." 5. The "I Review Papers in a Field I Don’t Understand" Dilemma User "Highlight" (a biochemist) was roped into reviewing ML theory: "I’m from a biochemistry background. They assigned me 5 ML papers. I’m scrambling to understand the math over the weekend. They must be desperate." 6. The "R is Not the Real Numbers" Debacle User "better" vented: "A reviewer complained: ‘What is 𝐑? You never defined it. It can’t possibly mean the real numbers!’ …What else would it be?!" 7. The "Dataset Police" Strike Again User "Reicholguk" faced this absurdity: "A reviewer demanded we test on ‘popular’ datasets like Cora/Citeseer, ignoring that we already used Amazon Computer and Roman Empire graphs (which are standard in our subfield). Is this reviewer an AI? Even AI wouldn’t be this clueless." 8. The "I’ll Just Give Random Scores" Strategy Many users reported suspiciously patterned scores: "877129391241": "One of my papers got no reviews at all (blank). Another got all 1s and 2s." "陈靖邦": "Got 4443 after an ICLR desk reject. Is this luck or a sign reviewers just clicked randomly?" Why This Matters These aren’t just funny fails, they reveal deep flaws in peer review: Overworked reviewers (5+ papers, no opt-out). Mismatched expertise (biochemists judging theory). Lazy/bad-faith reviews (no comments, citation demands). Systemic randomness (scores with no justification). As User "虚无" warned: "If ICML keeps this up, no serious researcher will want to submit or review." The Big Question Should top conferences like ICML: Cap reviewer workloads? Allow expertise-based opt-outs? Penalize low-effort reviews? What’s your take? Share your worst review horror stories below! (Sources: Zhihu users "877129391241", "虚无", "CpGD7", "陈靖邦", "Jane", "Highlight", "better", "Reicholguk", "写条知乎混日子". Original posts here.)

CCF Recommendation Conference Deadline Transferred from ccf-deadlines CCF-A October 13 - 16, 2025 Lotte Hotel World, Seoul, Republic of Korea ACM Symposium on Operating Systems Principles Deadline: Fri Apr 18th 2025 08:59:59 BST (2025-04-17 23:59:59 UTC-8) Website of SOSP CCF-B October 12-18, 2025 Singapore International Static Analysis Symposium Deadline: Mon May 5th 2025 12:59:59 BST (2025-05-04 23:59:59 UTC-12) Website of SAS CCF-A November 16 - November 20, 2025 Seoul, South Korea International Conference on Automated Software Engineering Deadline: Sat May 31st 2025 12:59:59 BST (2025-05-30 23:59:59 UTC-12) Website of ASE CCF-B October 12-18, 2025 Singapore International Conference on Function Programming Deadline: Fri Jun 13th 2025 12:59:59 BST (2025-06-12 23:59:59 UTC-12) Website of ICFP CCF-A April 12-18, 2026 Rio De Janeiro, Brazil International Conference on Software Engineering Deadline: Sat Jul 19th 2025 12:59:59 BST (2025-07-18 23:59:59 UTC-12) Website of ICSE

AI Review for https://arxiv.org/abs/2503.10617v2 Summary This paper introduces a new method, Compositional Subspace Representation Fine-tuning (CS-ReFT), designed to adapt large language models (LLMs) to multiple tasks while avoiding cross-task interference. The key innovation of CS-ReFT is its use of multiple orthonormal subspace transformations, each specialized for a distinct skill or task, and a lightweight router that composes these subspaces dynamically. The approach is evaluated on the AlpacaEval benchmark, where it demonstrates a high win rate compared to both larger models and other parameter-efficient fine-tuning methods. Soundness 2 Presentation 3 Contribution 2 Strengths Novel Method: Introduces CS-ReFT to address cross-task interference in multi-task learning using multiple orthonormal subspace transformations and a lightweight router. Unique Approach: Distinguishes itself from traditional weight-based methods like LoRA by applying representation editing. Clear Structure: The paper is well-organized and clearly written, with a logical flow from the introduction to method, experiments, and conclusions. Visual Aids: Uses mathematical formulations and diagrams to effectively explain the subspace transformations and router mechanism. Weaknesses Scalability Concerns: Does not sufficiently discuss how CS-ReFT scales, especially in scenarios with a large number of tasks, as maintaining separate low-rank transformations may become computationally and storage-intensive. Interpretability Issues: Lacks an analysis of the interpretability of the router's decisions. Insights into why specific weights are assigned to different subspaces could enhance trust in the model's behavior. Questions Scalability: How does CS-ReFT perform as the number of tasks increases? Are there strategies to optimize or reduce the computational overhead of maintaining separate subspace transformations for each task? Router Interpretability: Can the router's decisions be interpreted or visualized to understand why specific subspaces are activated for certain inputs? Are there any patterns that emerge which could provide insights into the model's behavior? Flag For Ethics Review No ethics review needed. Details Of Ethics Concerns (None) Rating 5 Confidence 4

The peer review process has long been upheld as the gatekeeper of quality, rigor, and integrity in academic discourse. Yet, as the recent NeurIPS 2024 keynote incident has reminded us, peer review does not extend its reach to one of the most visible stages in our field: the invited keynote. https://youtube.com/shorts/nYY4cIg-_8I?si=bkB964WIg7QllyKR In her keynote talk titled "How to optimize what matters most", MIT professor Rosalind Picard presented a slide that quoted a Chinese student, expelled from a top university for AI misuse, claiming, "Nobody at my school taught us morals or values". The slide included Picard’s commentary: "Most Chinese who I know are honest and morally upright". This statement triggered widespread backlash, with AI researchers such as Jiao Sun of Google DeepMind and Yuandong Tian of Meta calling out the racial undertones and singling out of nationality as discriminatory. Picard has since issued a public apology, acknowledging that referencing the student’s nationality was "unnecessary" and "caused unintended negative associations". NeurIPS organizers also released a statement clarifying that the comment violated their code of conduct and would be addressed internally. But this raises a deeper issue: How did such a racially loaded narrative make it into one of the most prestigious speaking slots in AI without any checks? In peer-reviewed papers, reviewers scrutinize phrasing, bias, and even the implications of an algorithm’s performance on diverse populations. Yet keynotes, arguably more visible and influential, operate outside the peer review system. A keynote speaker’s reputation is often deemed sufficient vetting. This incident, however, illustrates how even well-meaning researchers can unintentionally perpetuate harmful stereotypes when unchecked. What’s more concerning is the narrative framing: invoking a student’s nationality to make a moral point, while generalizing from a single anecdote to an entire culture. It underscores a blind spot we often fail to address in AI research: that ethical responsibility is not just about the fairness of our models, but also the fairness of our words and stories. If NeurIPS had subjected keynote slides or abstracts to a light peer review (or even a DEI advisory group), this incident might have been avoided. The community deserves keynote talks that not only inspire technically, but also model the inclusive values we preach in our papers. This is a wake-up call. If we expect AI models to mitigate bias, we must also ask whether our academic institutions and conferences are holding themselves to the same standard. Bias is not just a data problem; it’s a human problem. And no, peer review won’t solve everything, but perhaps it should go a bit further than we thought. Your thoughts? How can major conferences better safeguard against such missteps? Should keynotes be more rigorously reviewed, both ethically and technically?

Recently, someone surfaced (again) a method to query the decision status of a paper submission before the official release for ICME 2025. By sending requests to a specific API (https://cmt3.research.microsoft.com/api/odata/ICME2025/Submissions(Your_paper_id)) endpoint in the CMT system, one can see the submission status via a StatusId field, where 1 means pending, 2 indicates acceptance, and 3 indicates rejection. This trick is not limited to ICME 2025. It appears that the same method can be applied to several other conferences, including: IJCAI, ICME, ICASSP, IJCNN and ICMR. However, it is important to emphasize that using this technique violates the fairness and integrity of the peer-review process. Exploiting such a loophole undermines the confidentiality and impartiality that are essential to academic evaluations. This is a potential breach of academic ethics, and an official fix is needed to prevent abuse. Below is a simplified Python script that demonstrates how this status monitoring might work. Warning: This code is provided solely for educational purposes to illustrate the vulnerability. It should not be used to bypass proper review procedures. import requests import time import smtplib from email.mime.text import MIMEText from email.header import Header import logging # Configure logging logging.basicConfig( level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s', handlers=[ logging.FileHandler("submission_monitor.log"), logging.StreamHandler() ] ) # List of submission URLs to monitor (replace 'Your_paper_id' accordingly) SUBMISSION_URLS = [ "https://cmt3.research.microsoft.com/api/odata/ICME2025/Submissions(Your_paper_id)", "https://cmt3.research.microsoft.com/api/odata/ICME2025/Submissions(Your_paper_id)" ] # Email configuration (replace with your actual details) EMAIL_CONFIG = { "smtp_server": "smtp.qq.com", "smtp_port": 587, "sender": "your_email@example.com", "password": "your_email_password", "receiver": "recipient@example.com" } def get_status(url): """ Check the submission status from the provided URL. Returns the status ID and a success flag. """ try: headers = { 'User-Agent': 'Mozilla/5.0', 'Accept': 'application/json', 'Referer': 'https://cmt3.research.microsoft.com/ICME2025/', # Insert your cookie here after logging in to CMT 'Cookie': 'your_full_cookie' } response = requests.get(url, headers=headers, timeout=30) if response.status_code == 200: data = response.json() status_id = data.get("StatusId") logging.info(f"URL: {url}, StatusId: {status_id}") return status_id, True else: logging.error(f"Failed request. Status code: {response.status_code} for URL: {url}") return None, False except Exception as e: logging.error(f"Error while checking status for URL: {url} - {e}") return None, False def send_notification(subject, message): """ Send an email notification with the provided subject and message. """ try: msg = MIMEText(message, 'plain', 'utf-8') msg['Subject'] = Header(subject, 'utf-8') msg['From'] = EMAIL_CONFIG["sender"] msg['To'] = EMAIL_CONFIG["receiver"] server = smtplib.SMTP(EMAIL_CONFIG["smtp_server"], EMAIL_CONFIG["smtp_port"]) server.starttls() server.login(EMAIL_CONFIG["sender"], EMAIL_CONFIG["password"]) server.sendmail(EMAIL_CONFIG["sender"], [EMAIL_CONFIG["receiver"]], msg.as_string()) server.quit() logging.info(f"Email sent successfully: {subject}") return True except Exception as e: logging.error(f"Failed to send email: {e}") return False def monitor_submissions(): """ Monitor the status of submissions continuously. """ notified = set() logging.info("Starting submission monitoring...") while True: for url in SUBMISSION_URLS: if url in notified: continue status, success = get_status(url) if success and status is not None and status != 1: email_subject = f"Submission Update: {url}" email_message = f"New StatusId: {status}" if send_notification(email_subject, email_message): notified.add(url) logging.info(f"Notification sent for URL: {url} with StatusId: {status}") if all(url in notified for url in SUBMISSION_URLS): logging.info("All submission statuses updated. Ending monitoring.") break time.sleep(60) # Wait for 60 seconds before checking again if __name__ == "__main__": monitor_submissions() Parting thoughts While the discovery of this loophole may seem like an ingenious workaround, it is fundamentally unethical and a clear violation of the fairness expected in academic peer review. Exploiting such vulnerabilities not only compromises the integrity of the review process but also undermines the trust in scholarly communications. We recommend the CMT system administrators to implement an official fix to close this gap. The academic community should prioritize fairness and the preservation of rigorous, unbiased review standards over any short-term gains that might come from exploiting such flaws.

The ACL Rolling Review (ARR) process is undergoing significant changes in 2025 to address concerns around review workload, decision reliability, and consistency with conferences. Here’s a summary of the key updates: New Review Cycle: 5 Rounds Per Year Instead of monthly submissions, ARR will now have 5 review rounds per year. The planned review cycles for 2025 are February, May, July, October, and January 2026. This change aims to balance reviewer workload and submission volume while maintaining quality. Meta Reviews to Decide Findings ARR will now make Findings decisions directly instead of passing the responsibility to conferences. This is expected to improve alignment between ARR and conference decision-making. However, concerns remain about reliability and the potential impact on acceptance rates. Open Questions and Community Concerns 1️⃣ Reliability of ARR Meta Reviewers: Experiments at ACL 2025/EMNLP 2025 will assess decision consistency. 2️⃣ Acceptance Rate Changes: Current conferences accept ~35-40% of submissions. Higher acceptance via ARR might raise quality concerns. 3️⃣ Unified Findings Venue: Previously, Findings papers were associated with specific conferences, such as Findings of ACL, Findings of EMNLP, and Findings of NAACL. The proposed change would consolidate all Findings papers under a single umbrella venue (Findings of ACL). Potential benefits: Simplifies submission logistics and reduces ambiguity about which conference a Findings paper belongs to. Potential drawbacks: Findings papers would no longer be tied to prestigious main conferences (e.g., ACL, EMNLP), which could lower their perceived status and impact in the community. For example, in some academic institutions and funding agencies, having “Findings of ACL” may not be viewed as highly as “Findings of EMNLP” or “Findings of NAACL” due to differences in how venues are ranked. What’s Next? The 5-cycle review system is confirmed for 2025. The Meta Review-driven Findings decision is still under discussion, with potential refinements based on feedback. Researchers should stay updated on ACL policy announcements to navigate submission strategies effectively. Read more here: https://www.aclweb.org/portal/sites/default/files/COPR-5-cycle-report-to-publish.pdf What do you think of these changes? Will they improve the review process, or create new challenges? Let’s discuss! 

Recently, I was revising my paper and, out of curiosity, asked DeepSeek AI to give some sharp reviews of major AI conferences. The results were entertaining enough to share! [image: skynews-deepseek-app_6812411.jpg] CVPR The "crown jewel" of computer vision, with submission volume rivaling holiday train tickets and painfully low acceptance rates. CVPR is an event that raises the blood pressure of AI researchers globally. If your paper gets rejected, don't worry—90% of people share your fate. CVPR is like a battlefield for papers, with reviewers acting as ruthless supervisors holding magnifying glasses to pick faults. You're the student endlessly tormented. Hot topics? They change annually. This year it’s diffusion models; next year, it'll probably be something even more mystical. CVPR teaches us: forget eternity; pursue temporary insanity instead. NeurIPS The "Oscars" of machine learning, with acceptance rates heartbreakingly low. NeurIPS resembles an academic stand-up comedy show, where everyone gathers to share their "innovative ideas." But don’t celebrate prematurely; your paper might still be rejected for "inelegant experimental design." Reviewers here are like scholarly monks living deep in the mountains, meditating on how to perfect each paper. Their feedback sounds like verdicts from the void: "Sorry, your model isn't mystical enough." NeurIPS teaches one valuable lesson — before conquering the world, first conquer your reviewers. ACL The "ancestor" of NLP, acceptance rates low enough to shatter hearts. ACL is like a dating event for language models, where researchers showcase their masterpieces only to realize "other people's children" always seem better. Reviewers here act like literary inquisitors, scrutinizing your paper under a microscope for every potential logical flaw: "Why didn’t you use a larger dataset?" "Why didn't you try that model instead?" ACL reveals an essential truth: forget fairness; believe only in your capabilities. ICML The "Spring Festival Gala" of machine learning, enormous and influential. ICML is like an academic amusement park, where everyone chases the latest research trends. But remember, the entry ticket is your paper. Reviewers here act like academic police, checking if your paper meets their aesthetic standards: "Your experimental design isn't innovative enough," or "Your results aren't impressive enough." ICML reminds us there are no shortcuts, only relentless effort. ICLR The "rising star" of deep learning, with acceptance rates crushingly low. ICLR is an academic influencer’s hotspot, with everyone eager to showcase their research. However, the entry threshold is deceptively high. Reviewers here behave like fashion bloggers, critically evaluating if your paper aligns with their aesthetic tastes: "Your model isn’t flashy enough," "Your results aren't stunning enough." ICLR imparts the truth—work hard, but never forget to be trendy. The "mysticism" and "reality" of those conferences Submissions are lotteries: Acceptance often relies heavily on your luck. 🧩 Reviewers are puzzles: Their feedback can feel enigmatic, originating from mysterious realms. Academic rat race: Top conferences resemble massive "paper factories," with everyone desperately chasing trends and novelty, sometimes realizing too late that their efforts were in vain. Final thoughts If your goal is to publish in top conferences, brace yourself for a prolonged psychological battle. Yet remember, academic research aims to explore truths, not chase prestigious conference halos. View submissions as part of your scholarly journey and feedback from reviewers as opportunities for growth. True academic value isn't measured by how many top papers you publish, but by your contributions to advancing the field.