Shane Gu

Shixiang Shane Gu is a senior staff research scientist at Google DeepMind, working on Gemini Thinking. His research spans generative modeling, robot learning, and language-model reasoning. He co-developed Gumbel-Softmax, helped robots learn from shared experience, and co-authored zero-shot chain-of-thought prompting research that showed how a simple instruction could elicit reasoning without worked examples.

From learning algorithms to physical robots

Gu studied engineering science at the University of Toronto under Geoffrey Hinton’s supervision, then pursued doctoral research in machine learning at the University of Cambridge and the Max Planck Institute for Intelligent Systems. His doctoral supervisors were Richard E. Turner, Zoubin Ghahramani, and Bernhard Schölkopf. His research encompassed methods for training stochastic neural networks and algorithms for robotic control.

With Eric Jang and Ben Poole, Gu co-developed Gumbel-Softmax. Sampling a discrete choice—selecting one category from several—creates a difficulty for neural-network training because ordinary backpropagation cannot pass through that choice. Their method substitutes a smooth, differentiable sample that can gradually approach a categorical one. It provides a practical way to train models containing discrete latent variables using gradient-based optimization.

As a founding member of Google Brain’s robotics team, Gu worked on another expensive learning problem: physical experimentation. A robot cannot collect experience at the speed of a simulation, and unsuccessful attempts cost time. With Ethan Holly, Timothy Lillicrap, and Sergey Levine, he developed robotic manipulation with asynchronous off-policy updates. Multiple robots could contribute experience to a shared learning process, while the algorithm reused previous interactions. The team demonstrated a physical robot learning to open a door without prior demonstrations or a manually designed policy representation. Gu also led a dexterity moonshot, extending his work on learning to act in the physical world.

His subsequent work with Scott Fujimoto addressed learning entirely from existing data. Their 2021 minimalist approach to offline reinforcement learning combined an established reinforcement-learning algorithm with a behavior-cloning term and data normalization. The behavior-cloning term encourages the learned policy to stay close to actions represented in the dataset, reducing reliance on unreliable estimates for unfamiliar actions. The method offered a competitive baseline with relatively little additional machinery.

Eliciting and improving language-model reasoning

Gu shifted his main attention toward language models because he expected symbolic intelligence to advance faster than physical intelligence. At OpenAI, he worked as a senior researcher on the ChatGPT team and led Japan market entry. His GPT-4 contributions included model safety, the OpenAI Evals library, and model-graded evaluation infrastructure. At Google DeepMind, he subsequently led multilingual Gemini post-training before focusing on Gemini Thinking.

His language-model research examines how to elicit capabilities acquired during pretraining and how to develop them further:

• Reasoning without worked examples. With Takeshi Kojima, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa, Gu co-authored Large Language Models are Zero-Shot Reasoners. Adding “Let’s think step by step” improved performance across arithmetic, symbolic, and logical reasoning benchmarks without supplying task-specific worked examples. In his explanation of why natural-language reasoning works, intermediate steps give a model access to knowledge learned during pretraining; extracting capabilities through reinforcement learning can be computationally expensive.

• Self-generated reasoning as supervision. With Jiaxin Huang and colleagues, Gu investigated language-model self-improvement. A pretrained model generated explanations and answers for unlabeled questions. Agreement among sampled answers helped identify high-confidence solutions, which became fine-tuning targets. The experiments showed improvements on several reasoning benchmarks without ground-truth answer labels, using a selection procedure to determine which generated material became supervision.

• Physical simulation for grounded reasoning. With Ruibo Liu, Jason Wei, and colleagues, Gu co-authored Mind’s Eye. The system used the MuJoCo physics engine to simulate possible outcomes of a physical reasoning question, then supplied the results to a language model. It connected his interests in control and prediction with language-model reasoning: a model could receive information from a simulation when answering questions about the physical world.

Reasoning in space and time

Gu’s more recent research extends that inquiry to generative video. He co-authored Video models are zero-shot learners and reasoners, which tested Veo 3 on tasks beyond ordinary video production. Experiments included object segmentation, edge detection, physical understanding, tool-use simulation, and early forms of maze and symmetry solving. They examined whether a model trained to generate video also acquired useful capabilities for understanding and manipulating visual information.

For Gu, reasoning in space and time complements reasoning through language. Natural-language intermediate steps can draw on pretrained knowledge, but language cannot fully express motion, spatial relationships, or sensory detail. He sees cooperation between language and video models as a route toward richer physical prediction. Agentic workflows offer an incremental way to explore that cooperation, while reliable instruction following and reduced errors remain research goals.

His conception of world models is tied to the predictive model used in model-based reinforcement learning. This connects his multimodal research direction to his robotics background: video models could help anticipate physical consequences, extending their usefulness beyond producing convincing imagery.

Human judgment as a source of model improvement

Gu argues that prompting remains valuable because users notice differences that an automated system may overlook. He described a manga artist’s discomfort with subtly unnatural eye gaze in generated imagery: a detail that can escape general preference judgments may matter greatly to someone who has developed a trained eye. For Gu, refining that sensitivity and expressing the desired changes remain useful even as models become better at generating and revising their own prompts.

He also treats detecting semantic inconsistencies in generated video as an important intelligence task. A scene can look appealing while its events fail to make sense. Improving video understanding could help models recognize such failures and gradually reduce the human labeling work needed to evaluate them.

Gu extends this feedback loop to deployed applications. He uses a broad definition of post-training, encompassing the work between pretraining and the final user experience. In that view, engineers working with customers can improve how models are used, uncover failures in real tasks, and carry those insights back into model development. His discussion of customer feedback and forward deployed engineering links model improvement to knowledge that is difficult to recover from finished outputs alone: the decisions, iterations, and conversations behind a piece of work.