Gemini 4 Argon Safety & Security Report
Monitoring agent behavior and strengthening safeguards against misalignment in Gemini 4.
Li Ding 丁立
I’m a researcher at Google DeepMind working on AI control and agent safety.
I study how AI agents behave, where unexpected risks emerge, and how to make increasingly capable systems more trustworthy.
Selected research
All publicationsSecuring the future of AI agents
Research on monitoring and controlling increasingly capable AI agents.
Open-ended discovery with human feedback
Helping AI discover a wider range of useful and creative solutions through human feedback.
Learning from different human preferences
Learning from human feedback while accounting for the different contexts and preferences people bring.
Personalizing language models on-device
Efficient ways to adapt language models to individual users on their own devices.
A little background
Before AI control, I worked on pretraining for Gemma. I earned my PhD at UMass Amherst with Lee Spector, studying evolutionary computation, alignment (with Scott Niekum), and open-ended learning (with Jeff Clune and Joel Lehman).
At MIT, I worked on computer vision and autonomous vehicles. I also helped develop the MIT AI Podcast, now the Lex Fridman Podcast.
Open-source contributions google/gemma_pytorch facebookresearch/d2go pyribs MIT Deep Learning
Connect
I’m based in the San Francisco Bay Area and happy to chat!