Former ByteDance researcher Keyu Tian has come under renewed scrutiny after responding to the controversy around his internship period, only to face a new set of allegations tied to model training interference and research integrity.
Tian had recently returned to the spotlight after launching a world model-focused AI startup and securing nearly $30 million in funding. But the dispute linked to his time at ByteDance has resurfaced alongside that attention.
Tian says his misconduct was limited and not a long-running attack
In an interview published by Elsewhere on Oct. 12, Tian acknowledged that he had interfered with another intern’s training program. He rejected the description that he had carried out a sustained, months-long sabotage campaign, and denied maliciously damaging other researchers’ work.
According to Tian, the incident began because another intern was using GPU resources in an unreasonable way and even stopping other people’s programs, while the team failed to resolve the issue effectively. He said he used GPT to help write code and manually interfered with that person’s experiments about five times over roughly 10 days.
Tian said he never lacked GPU access and never touched other researchers’ experiments. He added that he was willing to take responsibility for violating rules, but did not accept being portrayed as someone who deliberately sabotaged colleagues’ research. He also said the misconduct did not affect anyone else’s research or papers, and denied using viruses or offensive security techniques.
Public statement says the conduct began in March 2024
A public statement titled "From Intern A’s Perspective" lays out a very different timeline. It alleges that the interference began as early as March 2024 and continued through July, with Tian repeatedly modifying AI model weights, PyTorch source code and distributed training procedures. The statement also claims he adjusted his methods while participating in the team’s own troubleshooting process.
According to the statement, Tian altered the T5 model weights used in the LlamaGen project in March 2024 by injecting random noise, which caused abnormal training behavior. The author said LlamaGen at the time was a small research project using no more than 100 GPUs and was not yet part of the later competition for large-scale video model compute.
The statement says the issue was resolved only after researchers re-downloaded the T5 model weights. It also includes a code snippet said to come from Tian’s internal GitLab commit history. The code allegedly added random noise to model weights and then restored the original norm, making the altered weights appear numerically normal in scale.
If the source of that code and its execution record are confirmed, the account would conflict sharply with Tian’s claim that he interfered only with one intern’s experiments over about 10 days. At this stage, though, the allegations still rest mainly on the statement from the other side, and the full logs and code attribution have not been independently verified.
Statement also accuses Tian of presenting PixArt output as VAR results
The statement raises a separate accusation involving research integrity. It says Tian was responsible for training a VAR text-to-image model with more than 1,000 GPUs and regularly reporting progress to supervisors.
It further claims that Tian deployed what he described as VAR training results to an internal Discord environment so colleagues could test image generation quality. After he came under investigation, the team re-examined the model and allegedly found that the system being tested was not the VAR T2I model he was training, but the open-source PixArt model.
The statement goes further, alleging that Tian built a mechanism into the inference pipeline so that no matter what prompt a user entered, the program would automatically prepend the word "City" before sending the request to PixArt. The author said this may have been intended to make direct prompt-by-prompt comparisons with the open-source model less obvious.
According to the statement, researchers had noticed that generated images often contained unexplained city building outlines, but could not identify the reason until they later inspected the inference flow and found the suspected modification. The statement also says the actual VAR T2I training had not reached the level of performance Tian had previously shown, forcing the team to reassess progress and continue development afterward.
The allegation carries extra weight because Tian’s current startup profile and fundraising appeal are tied in part to his past work in visual generative models. Earlier reports said he founded an AI lab focused on world models and raised nearly $30 million from investors including 5Y Capital and IDG Capital, despite having no public product and a team of only about 10 people. The post-money valuation was reported at $200 million.
Past research record and current fundraising are now under closer review
Tian previously won the NeurIPS 2024 Best Paper Award as first author for work centered on Visual Autoregressive Modeling, or VAR. He is now seeking to extend that line of work into AI foundation models designed to understand the physical world.
The public statement directly urged Tian’s current partners and investors to independently verify training records, model provenance and research progress, and to check whether his account of past events matches the facts.
Five researchers are named as potential witnesses
Alongside technical materials, the statement names five researchers who it says are willing to testify based on personal experience: Goku co-first authors Shoufa Chen and Chongjian Ge, Goku project member Ting-Che Lin, LlamaGen first author Peize Sun, and fellow intern Shilong Zhang.
The statement says those individuals confirmed the parts related to their own experience and would provide testimony if needed. Even so, the allegations remain claims from one side of the dispute, and no complete independent technical finding has been presented in the source material.
ByteDance lawsuit is still in the second-instance process
ByteDance said publicly in 2024 that an intern had maliciously interfered with model training for a research project and that the person involved had been dismissed. The company later sued Tian and initially sought more than RMB 8 million in damages.
In the Elsewhere interview, Tian said the court in the first instance set damages at about RMB 500,000. He also said ByteDance had appealed, and the case remains in the second-instance process.
Tian repeated in the interview that he admits violating internal rules and is willing to pay compensation, but believes many of the public accusations against him were exaggerated or fabricated. The newer claims, including the alleged long-term training interference and the use of PixArt in place of his own research output, have yet to be independently verified.
