Zhiwei Deng 邓志伟
I am a Research Scientist at Google DeepMind. Previously, I was a Postdoctoral Researcher in the Department of Computer Science at Princeton University, working with Olga Russakovsky and collaborating with Karthik Narasimhan. I received my Ph.D. from Simon Fraser University, advised by Greg Mori.
My research centers on building intelligent learning systems that reason and act. I am particularly drawn to (1) reparameterizing knowledge through the input–output space rather than model weights, (2) perceptual world models via self-supervised and generative learning, and (3) embodied and language agents that plan over long horizons.
- Gemini core (Gemini 2.5, 3.1, 3.5, 4.0)
- Gemini [x] Thinking (research leads)
- Gemini pretraining
Research
I/O Space Knowledge Representation
Storing learned knowledge in the input–output space via dataset distillation and influence-based data selection, drastically reducing the data needed for training.
Visual Perception & Generative Models
Building visual tokenizers via iterative grouping, and leveraging diffusion models for inference-time semantic editing and domain expansion.
Embodied & Language Agents
Topological navigation agents in vision-language environments, and anticipatory planning agents for long-horizon computer control.
Large-Scale Multimodal Systems
Scaling multimodal foundation models, with contributions to large-scale systems such as Gemini at Google DeepMind.
News
- Oct 2026Gemini 4 Argon.
- Sep 2026Two papers accepted to NeurIPS 2026: Reinforced Fast Weights via Next-Sequence Prediction (Spotlight) and DD-Ranking: Rethinking the Evaluation of Dataset Distillation (Oral, Evaluations & Datasets Track).
- Jul 2025Gemini 2.5 Technical Report released. [arXiv]
- Jan 2024Perceptual Group Tokenizer accepted to ICLR 2024.
Selected Publications
* equal contribution · Full list on the Publications page or Google Scholar.
-
Tech Report and Blogs -
Reinforced Fast Weights via Next-Sequence PredictionNeurIPS 2026Spotlight -
-
-
-
-
-
-
-
-
