CSPaper: peer review sidekick

Now the issue has been resolved. We will keep monitoring the review service.

Some users may experience an issue where cspaper.org and review.cspaper.org always redirects to forum.cspaper.org, making it impossible to use our tool. This problem does not affect everyone. It only happens if you visited cspaper.org during our recent domain change. ️ How to Fix You need to clear the cached redirect in Chrome. Normal cache/cookie clearing will not fix this. Please follow the steps in this video guide: : Screen Recording 2025-09-21 at 17.36.14.mp4 If this still does not resolve the problem, reply here and we’ll help further.

Official Announcement and Statistics On September 19, 2025, the NeurIPS community collectively held its breath as decisions were released. This year, 21,575 valid submissions were received for the Main Track, of which 5,290 papers were accepted — an overall acceptance rate of 24.52%. The breakdown of accepted papers is as follows: Posters: 4,525 Spotlights: 688 Orals: 77 Authors of accepted papers are required to present in person at one of two locations — San Diego or Mexico City — while Copenhagen offers an optional satellite poster session (EurIPS). The deadline for camera-ready submissions is October 23, 2025 (AoE). Yet behind the neat statistics lies a storm: despite strong reviewer support and unanimous positive scores, many papers were rejected due to venue capacity constraints. As one official standard phrased it, only work of “foundational and groundbreaking significance” would make it through. [image: 1758401507323-neurips-2026-decision-teaser.jpg] The Controversy: "Full Scores, Still Rejected" The community has exploded with frustration, especially around the phenomenon now dubbed “physical rejections” — papers with all positive (even maximal) reviewer scores being declined. According to leaks and social chatter, Senior Area Chairs (SACs) were instructed to reject already-approved papers simply to keep acceptance numbers within bounds. The consequence: strong, solid contributions — often with unanimous reviewer endorsement — were discarded at the final stage. [image: 1758401533429-sac-accept-standard-case.jpeg] This has left both authors and reviewers disillusioned. Reviewers who poured effort into fair evaluations saw their recommendations overturned, while authors with glowing reviews were pushed back into the exhausting cycle of “review → revise → resubmit → re-review”. Heated Reactions Across Platforms Across Chinese WeChat groups, Zhihu, Reddit, and Twitter/X, reactions spanned despair, anger, and irony: Doctoral students reported “7 submissions, 7 rejections,” comparing NeurIPS to a lottery. Review anomalies surfaced: authors described Area Chair comments that were factually wrong (e.g., citing “table errors” in papers without tables, or referring to “8 environments” when only 4 were present). Inconsistent outcomes: Some papers with average scores of 5/5/5/4 were rejected, while others with weaker profiles sneaked in. Cynical humor emerged: “Real skill isn’t doing research — it’s writing in a way that doesn’t make reviewers think too hard, because if they think, they might misunderstand.” One Chinese blogger summarized: “Directly rejecting even all-5-score papers just to control the rate is exaggerated. NeurIPS tried the dual-city venue model, but it still didn’t solve the bottleneck.” Decision Patterns: Scores vs. Outcomes The 2025 decisions were marked by inconsistency: papers with identical scores often faced opposite fates. Below is a consolidated view of reported cases across platforms (Zhihu, Reddit, Twitter/X, etc.) — capturing the full spectrum of anecdotes. Score Pattern Outcome Remarks / Reported Issues 5/5/5/4 Rejected AC introduced factually incorrect criticisms (“8 environments” when only 4 existed; claimed “table errors” though no tables). 5/4/4/4 Rejected Meta-review contradicted reviewers; cited errors not present in the paper. 5/5/5/2 Rejected AC dismissed reviewer support; authors suspected quota cuts. 5/5/4/3 Rejected Despite strong confidence (mostly 5s), AC added new issues post-rebuttal. 5/5/4/4, 5 reviewers Accepted (poster) Mixed reviews; one reviewer didn’t update scores, others raised post-rebuttal. 5/5/5/5 (all positive) Some rejected Cases of full scores declined due to “venue capacity”; sparked the “physical rejection” debate. 5/5/4/2 Rejected Minor points from reviewers reframed as “major flaws” by AC. 5/5/4/4/3 Accepted (poster) AC acknowledged rebuttal but outcome still borderline. 5552 Rejected Multiple reports of this score pattern rejected in both main track and DB track. 5554 Mixed: some posters, some rejections A few spotlight upgrades; others inexplicably rejected. 5544 Mixed: some spotlights, others rejected Widely cited as inconsistent; many authors baffled. 5543 Both accepted and rejected Example: some became posters, others turned down despite positive rebuttals. 5443 Both accepted and rejected Borderline batch heavily dependent on AC stance. 5442 Rejected Even when concerns resolved in rebuttal. 4444 Accepted Many reported “all 4s” surprisingly accepted. 4443 Mixed outcomes Some accepted, others rejected. 4433 Mostly rejected Borderline scores penalized. 4333 Accepted (poster) One report of borderline acceptance. 5533 Mostly rejected Authors called it the “graveyard zone.” 5433 Mostly rejected Similar pattern, AC-driven. 3335 Accepted Outlier case; noted on Reddit. 4544 → 5554 Accepted (spotlight) Rebuttal boosted scores. 2345 → 3455 Accepted (poster) Successful rebuttal turnaround. 2345 → 5557 (2024 precedent) Accepted Used as hopeful comparison by authors. 34455 → 44455 Accepted (poster) Multi-reviewer score boost after rebuttal. 5333 → 5533 Accepted (poster) ID ~26k; noted as “lottery win.” 4332 → 4444 Accepted (poster) One of the rare low-starting-score rescues. 55443 → 55444 Accepted (spotlight) Multi-modal identifiability paper. 5542 Rejected Despite AC promising “accept if revised,” outcome was negative. 5555 Accepted (spotlight) Some cases confirmed. 5556 Accepted (oral) Examples with IDs around 9k. 5655 Accepted (oral) Example from Reddit. 45555 Accepted (poster) Case shared from ID ~25k. 444 Accepted (poster) Noted as surprising acceptance despite simplicity. 6543 Rejected Reported confusion (“3 did not respond”). 6554 Accepted Example of strong paper getting through. Observation: High-score rejections (5/5/5/4, 5/5/5/2, 5544, 5552) were the most controversial. Mid-score bands (4444, 4443) had surprisingly better chances than some high-score papers. AC overrides and venue constraints trumped numerical averages. As one Reddit AC admitted: “Some rejections were not about flaws, but simply space constraints. It was a bloodbath.” Perspectives and Proposals The controversy has revived debates on how top-tier AI conferences should adapt to ballooning submissions: Adopting an ACL-style Findings track: Accept papers for archival recognition without guaranteeing a talk/poster slot. This balances fairness with logistical limits. Decoupling acceptance from physical venues: Many argue the artificial scarcity of poster space shouldn’t dictate what science gets recognized. Transparency in decision overrides: Authors want clearer documentation when SAC or PC overturns unanimous reviewer recommendations. Others are already planning to redirect rejections to ICLR 2026 or CVPR. Some predict ICLR 2026 will see a submission boom as displaced NeurIPS/AAAI works seek a new home. Tip: you can test your paper towards ICLR or CVPR via our AI review tool: https://cspaper.org/, where ICLR 2026 review agent has the knowledge about this "Perfect Storm" (We give our estimated submission to ICLR 2026 review agent) Lessons and Reflections Despite the frustration, senior researchers urge balance: Not all rejections imply low quality. Sometimes it’s the byproduct of constraints, not scientific merit. Writing clarity matters. Even obvious points must be explained plainly — reviewers may interpret differently otherwise. Luck plays a role. As one blogger put it: “Acceptance sometimes just needs a little bit of fortune. Don’t lose faith in your work.” For those accepted, congratulations — prepare for San Diego or Mexico City. For those rejected, the community’s consensus is clear: recycle, refine, and resubmit. The research journey continues. Final Word NeurIPS remains the flagship of machine learning, but NeurIPS 2025 has exposed deep tensions: between quality and quantity, between science and spectacle, between fairness and prestige. Whether the community embraces reforms like Findings, or continues down the path of selective “physical rejection,” will shape not just future conferences but the very ecosystem of AI research. For now, the 2025 decisions will be remembered as a year when “perfect scores weren’t enough.”

On September 17, 2025, the DeepSeek-R1 paper was officially published as a cover article in Nature, marking the first large language model (LLM) to undergo rigorous peer review and appear in a top-tier scientific journal. This milestone demonstrates not only DeepSeek’s technical achievements but also a broader shift in how AI research is being evaluated and recognized within the scientific community. Read the Nature paper here Key Highlights of the Publication Cover Recognition The DeepSeek-R1 study appeared on the cover of Nature, with the striking tagline “Self-Help: Reinforcement learning teaches AI model to improve itself.” This signals the importance the scientific community attaches to the work, particularly in the area of AI reasoning and reinforcement learning (RL). [image: 1758392066832-v2_9219473482fb4ebcae6c29f10c149f56@000000_oswg275067oswg1080oswg548_img_000.jpeg] A Model for Reasoning Tasks R1 is specifically designed for reasoning-intensive tasks such as mathematics and programming. Unlike traditional LLMs, it prioritizes logical inference over text prediction. Nature highlighted it as a cost-effective rival to expensive US-developed AI tools, with the added advantage of being an open-weight model freely available for download. On Hugging Face, R1 has already surpassed 10.9 million downloads, making it the most popular reasoning-oriented open-source LLM to date. Training Cost and Infrastructure The supplementary materials of the paper revealed for the first time the training cost of R1: Training R1 directly: ~ $294,000 USD Base LLM investment: ~ $6 million USD Comparison: Still far below the tens of millions typically invested by competitors. Training was conducted primarily on NVIDIA H800 GPUs, which are subject to US export restrictions since 2023 and cannot be sold to China. Despite this constraint, DeepSeek achieved competitive performance at a fraction of the cost. Peer Review and Revisions Did They Have to Revise the Paper? Yes. Despite being a landmark achievement, DeepSeek-R1 still underwent the standard peer-review process. Reviewers requested the removal of anthropomorphic language and asked for more technical details, especially regarding data types and safety measures. According to Ohio State University researcher Sun Huan, the process strengthened the validity and reliability of the results. Hugging Face engineer Lewis Tunstall called it a “very welcome precedent”, stressing that peer review is critical for transparency and risk evaluation in LLM research. This proves that even groundbreaking AI work cannot bypass the established standards of scientific rigor. Innovation: Pure Reinforcement Learning The core innovation of DeepSeek-R1 is its reliance on pure reinforcement learning (RL) rather than human-labeled reasoning datasets. The model learns by receiving rewards for correct answers, enabling it to develop self-verification strategies without explicit human guidance. Efficiency is enhanced through Group Relative Policy Optimization (GRPO), which allows the model to score and evaluate its own trial outputs without external algorithms. As a result, R1 has become a major inspiration for subsequent RL research in AI throughout 2025, shaping how reasoning-focused models are trained. Invitation or Self-Submission? One of the main questions was whether this paper was invited by Nature or self-submitted. While no official confirmation exists, analysts strongly suspect it was invited: The preprint version, released in January 2025, already received 3,598 citations and fueled an AI craze in China, including a hedge fund windfall for Qifan Quant through DeepSeek. Nature has a history of chasing high-impact, hot-topic papers. DeepSeek itself had little incentive to self-submit, given its prior success. Thus, the balance of evidence suggests that Nature invited the paper. Broader Impact DeepSeek-R1’s publication signifies more than academic prestige: It sets a precedent for peer-reviewed AI models, ensuring transparency and scientific credibility. It demonstrates that cost-efficient AI development is possible, even under geopolitical constraints. It shows how open-source models can drive global adoption and innovation. Conclusion DeepSeek-R1’s appearance in Nature is a defining moment for AI research. It bridges the gap between industrial innovation and scientific recognition, proving that large language models can meet the highest academic standards. The work also highlights the growing importance of reasoning, reinforcement learning, and cost-efficient AI in shaping the next generation of intelligent systems. Full paper: DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning

Impl. based on nr0034je9.zip . Table A: Model Performance on NLP Benchmarks Model SST-2 (Acc) MNLI (Acc) QNLI (Acc) CoLA (Matthews) Avg Score BERT-Base 91.2 84.6 90.1 58.2 81.0 RoBERTa-Base 92.3 87.4 91.8 63.1 83.7 GPT-3 (175B) 94.1 88.9 93.0 66.4 85.6 Our Method 94.8 89.7 93.5 68.9 86.7 Table B: Ablation Study on Model Components (Evaluated on MNLI) Configuration Attention Mechanism Pretraining Corpus MNLI (Acc) Full Model Multi-head Self-Attn Custom + Public 89.7 – w/o Custom Corpus Multi-head Self-Attn Public Only 87.1 – w/o Attention Refinement Block Basic Self-Attn Custom + Public 86.5 – w/o Positional Embeddings Multi-head Self-Attn Custom + Public 85.2 – Random Initialization — — 72.4