Xiaolei Chu

Hi, I’m Xiaolei, a fifth year Ph.D. candidate at UC Berkeley, advised by Prof. Ziqi Wang.

My research focuses on LLM post training, generative retrieval, diffusion/flow models, and reinforcement learning. I also have a background in mechanics, uncertainty quantification, and statistical physics. I am always happy to connect and discuss potential collaborations.

Email / CV / Scholar / Github / LinkedIn

News

May 2026Joining Google DeepMind & YouTube as an intern, leading research in generative retrieval.
Mar 2026New preprint “Social Amplification Dominates Collective Hazard Response” is now on arXiv. The paper uses mean-field theory from statistical physics to explain how social networks amplify collective panic when communities face hazards, revealing the conditions under which individual fear cascades into a system-wide phase transition.

Selected Publications

	Shortcut or Semantics? Understanding Conditioning Interfaces for Steerable Generative Retrieval Xiaolei Chu, Shao-Chuan Wang, Yu-Neng Chuang, Zhaoheng Zheng, Sanjay S. Girija, Ruining He, Aniruddh Nath, Li Wei, Lichan Hong, Xinyang Yi Google DeepMind & YouTube, Aug. 2026 paper SteerGR studies how natural-language preferences steer Gemma 3 to generate semantic IDs for retrieval. Comparing joint conditioning with modular language-state injection reveals how conditioning interfaces affect retrieval accuracy, phrasing robustness, and retention of general reasoning ability.
	GeoScout: A Generalizable Geometry-Aware Policy for Active 3D Reconstruction Xiaolei Chu, Guanren Zhou, Ziqi Wang, Khalid M. Mosalam coming soon, 2026 code GeoScout conditions a reinforcement-learning next-best-view policy on refined geometry captions, using language as a prior over unobserved shape to reconstruct unseen objects with fewer views than a matched no-caption policy.
	Entity-Guided Expert-Grounded Visual Diagnosis for Structures in the Wild Guanren Zhou, Xiaolei Chu, Khalid M. Mosalam coming soon, 2026 Hierarchical SAM 3 retrieval and DINOv3 specialist classifiers ground a frozen vision-language model’s fine-grained diagnoses in segmentation masks, enabling traceability without new task-specific annotations or model fine-tuning.
	Ground Motion Generation via Latent Space Diffusion Xiaolei Chu, Maijia Su, Ziqi Wang coming soon, 2026 code A hierarchical generative framework couples a conditional diffusion prior with latent space diffusion for three-component ground motion synthesis. The prior learns the joint distribution of energy and duration conditioned on magnitude, distance, and site properties, while a condition-aware diffusion U-Net synthesizes spectrograms in a learned latent space. This probabilistic factorization captures energy–duration dependence and sample diversity while recovering expected magnitude scaling and distance attenuation.
	Social Amplification Dominates Collective Hazard Response Xiaolei Chu, Guanren Zhou, Marco Broccardo, Didier Sornette, Khalid M. Mosalam, Ziqi Wang arXiv preprint, Mar. 2026 arXiv Fine-tuned BERTweet estimates state-level stress prevalence from social-media posts to calibrate an interpretable network model, revealing that social influence outweighed direct exposure in over 80% of U.S. states.

Website source from Jon Barron.