Keyu Tian faces fresh allegations over ByteDance training sabotage and PixArt model misrepresentation
Former ByteDance researcher Keyu Tian is facing a new round of allegations after publicly responding to the controversy that led Chinese internet users to label him an "AI poisoning intern." In an interview published by Elsewhere on Oct. 12, Tian admitted he had interfered with another intern’s training jobs, but said the conduct was limited to roughly five manual disruptions over about 10 days and was tied to a dispute over GPU usage. He denied carrying out a months-long campaign or affecting other researchers’ work. A newly circulated public statement titled "From Intern A’s Perspective" presents a sharply different account. It alleges that from March to July 2024, Tian repeatedly altered model weights, PyTorch source code and distributed training procedures, and even adjusted his methods while taking part in internal troubleshooting. The statement also raises a separate research-integrity accusation, claiming Tian used the open-source PixArt model while presenting the output as results from a VAR text-to-image model he was supposed to be training. The statement includes partial code said to be linked to the incidents and names five researchers who are willing to testify based on personal experience. Still, the claims remain largely based on one side’s account, and the full operational logs and code ownership have not been independently verified.
