The report, titled” Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction”, presents a novel method for creating AI-generated pictures that Tian and four coauthors—all affiliated with either ByteDance or Peking University—claim is faster and more effective than its predecessors. The committee for NeurIPS Best Paper Award stated in a statement that” the overall quality of the paper presentation, experimental validation, and insights (scaling laws ) provide compelling reasons to experiment with this model.”
The council’s decision to award Tian the honor quickly became the subject of more extensive discussions online about how NeurIPS is run and the manner top AI researchers assess the job of their coworkers. ByteDance reportedly sued for over$ 1 million in restitution last month. The specifics of a incident that had been brewing on Chinese social media for weeks suddenly appeared on the English-language computer as a result of the information.
” NeurIPS gave best paper award to a super problematic work ( not the first time this has ever happened btw ),” Abeba Birhane, head of Trinity College’s newly established AI Accountability Lab, wrote on Bluesky. You’d think a conference would do due diligence before awarding a paper that directly conflicts with their values, if it prides itself on upholding the highest scientific andamp, ethical standards.
The paper received the honor, according to a NeurIPS spokesperson, not Tian himself. They directed WIRED to a section of the award committee’s statement that explained how the conference evaluated paper submissions. According to the search committees,” the search committees took all accepted NeurIPS papers into consideration equally and made decisions on their own based on the papers ‘ scientific merit, without taking into account any differences in authorship or other factors, in accordance with the NeurIPS blind review process.”
On Bluesky, Birhane and other AI researchers posted an anonymous GitHub blog post that was also published on HackerNews, Reddit, and other platforms in recent days, urging the academic AI community to reconsider awarding Tian the Best Paper award because of his” serious misconduct,” which it claims “fundamentally undermines the core values of integrity and trust on which our academic community is built.”
A list of technical strategies Tian allegedly employed to use ByteDance’s computing resources for his own work is included in the blog post. It accuses him of having “deliberately disrupted experiments, causing erroneous and irreproducible results. This caused the researchers to question the scientific accuracy of their findings, severely deteriorating both physically and mentally.
Additionally, it requested that the research be withdrawn from ByteDance in a “honorary gesture of respect for the company’s other researchers and the entire academic community.” A request for comment from WIRED was not immediately returned; neither did Jiang Yi, a generative AI research leader at ByteDance, who was listed as the “project lead” on the award-winning paper, or the other coauthors. A comment request sent to a school email associated with him did not arrive in Tain’s response.
The Tian accusations in the blog post are similar to those in a different, earlier GitHub blog that was written in Chinese on October 18. On Chinese social media at the time, there were rumors about a ByteDance intern who allegedly caused the tech giant millions of dollars in losses by thwarting the company’s ongoing AI experiments.
ByteDance attempted to amend the record on October 19 when it published a statement on Toutiao, a Chinese news aggregator. It claimed that an intern on the” commercial technology team” had been fired in August for” serious disciplinary violations,” but that some details of the story had been exaggerated in media reports, including that 8, 000 graphics processing units ( GPUs ), which are required to train high-powered AI models. ByteDance added that it had reported the intern’s behavior to his university.
The controversy highlights the lengths some AI researchers are willing to go in order to get access to a sizable number of GPUs that are on the market. Due to US export restrictions, which restrict the sale of numerous high-end semiconductors to China, supplies are especially limited there. Tian and his coauthors, however, are optimistic that their future research in areas like AI-generated video could help make running models less resource-intensive.
It is becoming “prohibitively expensive” to generate high-resolution videos with traditional AI models, the authors noted in their paper, adding they believe their methodology could provide a solution. We therefore anticipate a bright future for the use of VAR models in the field of video generation.
Zeyi Yang contributed reporting.
