Author丨Deng Zhemin
Editor丨Qi Chengyong
Recently, a suspected "academic fraud" incident caused an uproar in the domestic academic circle.
The eye of the storm comes from the just-concluded top academic conference ACM MM 2026. A Best Paper at this top international multimedia conference has been pointed out many doubts, questioning its possible fraud and the gold content of the title of the best paper in the international top conference.
AI Technology Review (Leifeng.com (public account: Leifeng.com)) sorted out the ins and outs of the matter and exclusively contacted the corresponding author of this paper.
01
origin
On November 14, ACM MM 2026 concluded in Brazil. Due to the arrival of the AI era, this conference has attracted much attention from the academic community.
The protagonist of this mystery is one of the 17 best papers, titled "Clifford-Algebra-Based Geometric Feature Fusion for Real-Time UHD Low-Light Image Enhancement".
This is a work on how to use AI to make images brighter. The direction is very popular and many scholars have been working in this field for decades. What is surprising is that the first author of the paper, Xiaohan Wang, is a second-year undergraduate student at Xi'an Electronic University.
A sophomore's first paper is best if he gets the top prize.
Scroll down the list page, and in the corresponding author column, there is no university or research institute after the name, but Independent, independent researcher.
In the context of academic circles, this signature usually corresponds to two types of people. One type does not require an organization. It is successful and famous, and the title is already louder than the organization. The other is without units.
It was originally an inspirational narrative about an independent researcher leading a talented girl to devote herself to research, bravely enter the top international conferences, and win the Best Paper. However, because of what happened next, it turned into a public opinion storm about academic fraud.
02
paper
According to the AI Technology Review survey, the earliest question appeared on November 14. Someone asked on Zhihu how to evaluate this Best Paper.
The questioner claimed to have done research in the field of low level vision for many years and was the first to reply to his own question, listing 5 questions.
First, the comparison method is deliberately suppressed. The paper compares old methods that have not been trained on the data set with methods that have been trained by themselves.
Second, describe the synthetic data as real shots. The paper states that the test data are photos taken in the real world, but the author of the original data set has written that the data is computer-generated.
Third, the strong ones did not enter the competition. The methods that are better than mine are only mentioned by name and not included in the comparison table; the ones that are included are unsupervised methods. The paper's own score is 20.56 points, which cannot even beat an old method in 2018 (21.67 points).
Fourth, suspicious users score. Attached to the paper is a user study. All 48 sub-items neatly end in .x0 or . The probability of this distribution occurring naturally is 1 in 3.55 × 10³³.
Fifth, the core selling point is not worthy of its name. The title of the paper seems to be based on the high-number beacon "Clifford Algebra", but if you break down the formula in the paper, it is actually the most common operation of "multiplying two numbers and then summing them", which has little to do with advanced mathematics.
This post quickly attracted the attention of researchers in the field, and the corresponding author also responded personally, calling the doubters to "boldly reproduce".
The back-and-forth academic dialogue attracted more attention.
Soon, another respondent (a master's student in electronic information at Tsinghua University) tried to reproduce it. He recognized the engineering route of the paper, but raised more comprehensive questions about the research motivation and baseline comparison data, pointing out that the academic training of the first work was insufficient, the corresponding author's guidance was not in place, and the conference review was not strict.
Excerpted from the conclusion of the respondent on Zhihu. For the complete answer, see: https://www.zhihu.com/question/2104933122630476610
Since the paper is not open source, doubts surrounding the data cannot be confirmed or falsified. The two sides are shouting at each other through a layer of frosted glass.
On November 17, someone turned their attention from the experimental part to the references: the 12th citation of the paper pointed to a non-existent paper; the 47th and 51st citations, the author information did not match the actual literature.
The three errors are distributed in different positions in the list and are of different types. They are unlikely to be "hand slips".
For readers who are not familiar with academic ethics, citing hallucinations is serious academic misconduct. The 12th fictitious reference alone is enough to constitute a retraction-level consequence, because in nature, it means that the author has implanted false support in the argument chain of the paper that cannot be traced.
Discussions around the technical details of the paper require a threshold, and citation issues are much more common. As the allegation was proven, the discussion escalated.
03
author
Questions about the paper quickly turned to people.
The first is Xiaohan Wang, a second-year undergraduate student at Xi’an University of Electronic Science and Technology. This is the only publicly available paper. In this field, a teacher who has graduated with a doctorate for many years may not be able to get the Best Paper throughout his career, but this sophomore student got it for his first paper.
More discussions focused on the corresponding author, Zheng Zhuoran. The author's information did not fill in the university or research institute where he was located, but "Independent". There is nothing wrong with publishing a paper as an independent researcher as the corresponding author. However, someone released a screenshot. In August this year, Zheng posted on Xiaohongshu, offering clearly priced thesis guidance services.
Publicly providing paid services for specific research directions and promising to "guarantee quality and quantity" has gone beyond the scope of normal academic exchanges and has an obvious nature of commercial transactions. "5,000 before winning, 5,000 after winning", implying that the service price is directly linked to whether the paper can be accepted. This is very much like a risk agency model, where the final payment is paid after the paper is successfully published.
Highlighted "full guidance" is also in a fuzzy zone. If the essence is to complete core work for students (such as proposing ideas, writing code, revising papers to a publishable state), then this crosses the line into the category of ghostwriting. Even if it is just providing ideas and modifications, but combined with the premise of payment, its nature can easily slip into academic misconduct.
One stone stirred up a thousand waves, and the discussion on Zheng Zhuoran's identity became more and more intense on the Internet. Some people suspect that he took on private jobs at a medical research institution and helped students get rejected by supervising their thesis. Some people claiming to be close to him said that he once taught at Qilu University of Technology but was expelled for making false resumes. He now lies to students that he is from the National University of Defense Technology.
Public information shows that Dr. Zheng Zhuoran graduated from Nanjing University of Science and Technology, worked as a postdoctoral fellow at Sun Yat-sen University, and then joined Qilu University of Technology.
There is no public information available as to whether his current teaching status will continue. On his academic homepage (Open Review and Research Gate), his affiliation is shown as Nanjing University of Science and Technology, his Zhihu profile reads "Doctorate in Computer Science, focusing on low-level visual tasks and representation learning", and his Xiaohongshu profile reads "Computer Postdoctoral Fellow (Sun Yat-sen University), teacher in the Department of Computer Science, Computer Vision", neither of which explicitly mentions the teaching position.
Zheng Zhuoran’s Zhihu account homepage
AI Technology Review (Leifeng.com) noticed that Zheng Zhuoran's latest achievement (submitted to arXiv on November 3) has a total of eight pages and three references. There are two people in the author column, one is him, the other is signed "DouBao", and ByteDance's AI assistant Dou Bao, the pinyin is correct.
On the morning of the 18th, Zheng Zhuoran publicly stated: "Because my supervision was not in place, we decided to withdraw the manuscript." Subsequently, he successively deleted many replies on Xiaohongshu, Zhihu and other platforms, and cleared the contents of Xiaohongshu's homepage. At present, regarding this controversy, the only reply he has left on the public platform is to announce the withdrawal of the article.
Someone discovered that Zheng Zhuoran's postdoctoral supervisor Ren Wenqi suddenly removed him from his research group list.
04
respond
On the evening of the 18th, Zheng Zhuoran was interviewed by AI Technology Review, and his voice sounded a little tired.
Zheng Zhuoran clarified the content on the Internet one by one. First of all, the reason why it is labeled Independent (independent researcher) is because during the period of directing this work, Zheng resigned from Qilu University of Technology and did not find a job, so he had no unit to apply for.
It was reported on the Internet that he was "fired for fabricating his resume", but he firmly denied it: the information can be found online, including which papers he has published, how could he falsify it? Zheng Zhuoran said that the real situation is that after he left Qilu University of Technology, he took a gap for a while and then entered the National University of Defense Technology.
Talking about the doubts on the Internet that his supervision of the thesis involved an exchange of interests and an improper process, his answer was that there was no exchange of interest and he guided this work out of friendship. As for the so-called joining of thesis mills, Zheng’s original words were: “I can swear that I have not joined any thesis mills or organizations, even though they have come to me many times.”
At the same time, Zheng mentioned that his original intention of doing scientific research has always been very pure. He and his students are very limited in terms of hardware. The cards are rented from Xianyu. When the server expires, the original data will be lost, so he has always pushed students. The rhythm of group meetings is once every few days.
Zheng mentioned that some people attacked that his undergraduate degree was only from Zaozhuang University. In this regard, he did not shy away from his undergraduate degree and believed that it did not affect his pursuit of scientific research.
From submission to Best Paper, a paper has to go through four gates: reviewer scoring, AC (field chair) making the acceptance judgment, the Program Committee’s final review, and finally the Awards Committee selecting the Best Paper. Some netizens said that the paper was clearly rejected during the AC review stage, and three reviewers gave it a perfect score of 5, implying that the paper team had collusion with the conference party.
Zheng’s version is a completely different story: the initial score of the paper was in the middle of the pack, but it was the revised draft that turned the situation around; in the middle of this year, when he received the news of the successful draft, the AC’s opinion was to recommend the poster; a month or two later, he received an email that the paper was rated as oral; and the day before the award, he received a text message that the paper was selected for Best Paper.
System interface for collecting papers, provided by Zheng Zhuoran
The key to the modification is to highlight the application value: this method can be deployed on Android mobile phones and has commercial prospects, while the multimedia application track focuses on implementation. After the revision, the three reviewers unanimously gave it a score of 5.
There was no fraud in the algorithm, which was what he wanted to clarify most. Zheng Zhuoran cited the reproduction results of Tsinghua University's master's degree on Zhihu. He said that the paper basically has no problems in terms of performance and parameter quantity. The awards are selected by a foreign-led jury, and Zheng Zhuoran has no control over it. "If we really had it done, we wouldn't have found an acquaintance in Brazil to come on stage to accept the award for us."
When asked about hallucinatory citations, Zheng Zhuoran said that he had many students and could not check every document in each manuscript one by one. He reminded the students countless times, but something still happened. This is his first submission and he lacks experience. He believes that the main responsibility lies with himself, and plans to suspend scientific research work and start again after a year.
In addition, Zheng hopes to convey three messages to the public through us. First, he is considering entering the industry and hopes to do some new work. At the same time, "I hope the academic community will give Yi Zuo another chance." Finally, he emphasized that "Xidian is a good school."
Regarding this controversial paper, all the authors of the team have unanimously agreed to withdraw the paper. They have sent a joint email to the conference chair's email address requesting the paper be withdrawn, and mentioned that "if no reply is received within two days, we will urge you to expedite it again."
05
Review
According to AI Technology Review, the award structure and review ecology of ACM MM 2026 provide some more structural explanations for why this paper was able to pass through the four review gates and reach the highest podium of the main conference.
In 2024 and 2025, the ACM MM main conference will only award one Best Paper, and there will also be a Best Student Paper. In December 2025, the ACM MM Steering Committee passed new rules. Taking into account the increase in the number of paper submissions, in 2026, about 5% of the Outstanding Papers will be selected from the accepted papers according to the theme, and then the Best Paper, which will account for about 1% of the total number of accepted papers, will be selected.
Excerpted from "Report of the December 2025 vote of the ACM Multimedia Steering Committee", hereinafter referred to as "Award Guide"
This year, the main conference of ACM MM accepted a total of 1,703 papers and awarded 17 Best Papers, which is in line with the officially announced selection ratio.
However, there were as many as 17 "best" papers in an academic conference, which is still beyond people's imagination of top papers.
In the past two years, the number of submissions to top conferences in the AI field has been rising. ICLR's annual submissions have increased from 1,013 articles in 2018 to nearly 20,000 this year, an increase of nearly 18 times in eight years. The number of submissions has skyrocketed, but reviewers can’t keep up. A researcher who has served as the field chair of a top conference said that in the past, a reviewer's workload was two or three articles, but now it is often ten.
The quality of review is visibly declining. This year, ICML conducted a full public replication audit for the first time. Of the 168 oral presentation papers, only 8, or less than 5%, could replicate more than 80% of the conclusions. Hallucinated citations are not a unique problem in this controversial paper. Just this year, AI-generated papers with fictitious citations also appeared at ICLR.
Quoting a document that does not exist does not require professional ability to verify, but it has passed through the reviewers, AC, program committee, and awards committee, all the way, and finally hung up on the public version of a Best Paper.
Rather than blaming the failure of a single reviewer, it is better to say that this is a systematic failure of the review system under the flood of submissions.
AI Technology Review noted that the "Award Guidelines" after last year's reform have made stricter avoidance of the conference chairperson and PC to ensure the independence of the awards.
The AwC's responsibility is to "discuss and decide on the Best Paper", and the document does not include a provision for "checking the authenticity of citations in candidate papers". This means that if citation issues are not blocked at the review stage, there is theoretically no mechanism to block them at the award stage.
In addition to the "non-verification" in the process, what deserves attention is the actual carrying capacity of the award process. 17 Best Papers, perhaps the other side of loss. From the only best to one for each track, it seems to recognize more good work, but the actual situation may be: when the best can be distributed in batches, its gold content is diluted, and some verifications that should be more stringent will eventually become superficial.
Currently, regarding the controversial paper, ACM MM 2026 Program Co-Chair Yan Junchi confirmed to AI Technology Review that the Bureau has received the retraction application email and will decide whether to make an official public response in the near future.
Will this turmoil caused by the "sophomore student + independent researcher" winning the best award at the conference end with the retraction of a paper, or will it become an opportunity for the conference to reform its review and award mechanism? AI technology reviews will continue to pay attention.
NeurIPS 2026 is about to be held. We welcome participating researchers, students, and readers who pay attention to the cutting-edge progress of AI to join our communication group.
Here, you can meet peers, exchange research, share enrollment information, scientific research cooperation, make dinner appointments, submit papers, form teams to attend conferences, etc., and find potential partners.
During the conference, we will also continue to bring on-site reports, guest speech interpretations, cutting-edge trends and forum highlights.
See you at the venue and in the group.
? Join the group portal: Scan the QR code to join the group or add WeChat szhsn610, remarks: NeurIPS+unit/school+name+direction.
Original articles from Leifeng.com, unauthorized reproduction is prohibited. See the reprint instructions for details.
edit
Resend Email Modify Email




