Shane Gu

Shane Gu, also known as Shixiang Shane Gu and publishing in earlier work as Shixiang Gu, is an AI researcher whose contributions include zero-shot chain-of-thought prompting and training language models on their own generated reasoning. As of September 2026, his personal biography lists him as a Senior Staff Research Scientist at Google DeepMind focusing on Gemini Thinking.[2][3][4][5]

In 2022, Gu coauthored Large Language Models are Zero-Shot Reasoners with Takeshi Kojima, Machel Reid, Yutaka Matsuo and Yusuke Iwasawa. Their zero-shot chain-of-thought method used a single prompt—“Let's think step by step”—without handcrafted examples of how to solve each task. The paper, accepted to NeurIPS 2022, reported that this approach increased text-davinci-002 accuracy from 17.7% to 78.7% on MultiArith and from 10.4% to 40.7% on GSM8K. These were results for the named model and benchmarks, demonstrating how prompting could elicit reasoning capabilities without task-specific examples.[4]

Later that year, Gu joined Jiaxin Huang, Le Hou, Yuexin Wu, Xuezhi Wang, Hongkun Yu and Jiawei Han on Large Language Models Can Self-Improve. The team used chain-of-thought prompting and self-consistency to generate high-confidence answers with rationales for unlabeled questions, then fine-tuned a pretrained model on those solutions. They reported GSM8K accuracy rising from 74.4% to 82.1% for a 540-billion-parameter model, with ablations indicating that training on the reasoning was critical. The procedure required no ground-truth labels for this self-improvement stage; it began with an already pretrained model.[5]

Gu's personal biography describes earlier roles leading Gemini Post-Training's Multilinguality team, serving as a senior researcher on OpenAI's ChatGPT team and leading OpenAI's Japan market entry. It also identifies him as a founding member of Google Brain Robotics. In the publisher's account of his 2026 generative-media panel, Gu explained his move from robotics and dexterity toward language models through his expectation that symbolic intelligence would advance faster than physical intelligence.[2][8]

At the AI Engineer World's Fair 2026 generative-media panel, alongside Dumitru Erhan and Nicole Brichtova, Gu advocated combining language reasoning with video models. He argued that natural-language reasoning can draw on knowledge acquired during pretraining, while video supplies spatial and temporal information that language alone does not adequately represent. He described agentic workflows connecting Gemini understanding with Omni as an area of exploration, while treating eventual consolidation into one model as a forecast.[8]

Gu also argued for preserving human sensitivity as models automate more generation and prompting. He recalled a manga artist finding subtly incorrect eye gaze unsettling, illustrating how expert judgment can detect failures in otherwise polished images. In Richard MacManus's July 2, 2026 account, Gu urged people to notice outputs that remain wrong, generic or insufficient: “Always find your sensitivity.”[8][1]

For Gu, customer feedback should improve both the surrounding software and the model itself. In the panel's publisher account, he described forward deployed engineering as a two-way relationship: users can improve the harness through which they use a model, while researchers take the same observations upstream into model development. He also advocated turning human judgments about generated-video errors into better machine understanding, so evaluation effort could become reusable.[8]

Before his language-model work, Gu studied how to train stochastic neural networks and make reinforcement learning practical on physical robots. He earned an Engineering Science bachelor's degree at the University of Toronto under Geoffrey Hinton's supervision, followed by a machine-learning PhD at Cambridge and the Max Planck Institute for Intelligent Systems, supervised by Richard E. Turner, Zoubin Ghahramani and Bernhard Schölkopf. A historical Creative Destruction Lab profile also records a Google internship hosted by Ilya Sutskever and Sergey Levine.[2][3]

With Eric Jang and Ben Poole, Gu coauthored Categorical Reparameterization with Gumbel-Softmax, first submitted in November 2016 and listed among his ICLR 2017 publications. The method addresses the difficulty of backpropagating through discrete categorical samples: it replaces a nondifferentiable sample with a differentiable Gumbel-Softmax sample that can be smoothly annealed toward a categorical distribution. This provides a gradient estimator for training models with categorical latent variables.[6][2]

Gu and Ethan Holly were equal contributors to Deep Reinforcement Learning for Robotic Manipulation with Asynchronous Off-Policy Updates, coauthored with Timothy Lillicrap and Sergey Levine, first submitted in October 2016 and presented at ICRA 2017. The team trained deep Q-functions off-policy and reduced training time by letting multiple robots pool policy updates asynchronously. Their experiments covered 3D manipulation in simulation and a door-opening skill on real robots, learned without prior demonstrations or manually designed representations.[7][2]