Xinru Liu

665 Commonwealth Ave (314) · Boston, MA 02215 · (508) 369-1863 · liu [dot] xinruuu [at] gmail [dot] com

I am a Statistics Ph.D. candidate at Boston University supervised by Yves Atchade. My current research focuses on both theoretical and computational aspects of high-dimensional Bayesian statistics with applications in sparse deep learning and reinforcement learning. My currect ongoing research topics include inverse problems, thompson sampling and soft deep-Q network. I obtained my B.A in Mathematics and Computer Science, minor in music performance from Wheaton College (MA). In my leisure time, I love going to classical concerts, playing the piano, and doing outdoor activities.


Work Experience

Applied Science Intern

Liberty Mutual Insurance Solaria Labs
June 2023 - August 2023
  • Background: Enhanced underwriting decisions by fine-tuning LLM using Human In The Loop approach.
  • Improved the Flan-T5 LLMs’ performance in classifying risk characteristic queries (10% ↑ in F1-score), using prompt engineering and the Reinforcement Learning from Human Feedback fine-tuning algorithm(PPO).
  • Reduced the GPU memory requirement and computational cost by integrating LoRA reparameterization technique.
  • Facilitated reproducibility among team members by maintaining a GitHub repo adhering to industry best practices.


Graduate Data Science Intern

Liberty Mutual Insurance Office of Data Science
June 2022 - August 2022
  • Background: Identified similar business descriptions across diverse economic sectors using Hierarchy in NAICS Code.
  • Approach 1: Improved the baseline accuracy of semantic search for code prediction by 4% by developing a customized hierarchical similarity metric on MPNet embeddings through Breadth-first search (BFS) algorithm.
  • Approach 2: Boosted classification accuracy from 87% to 95% by engineering a hierarchical NLP classifier that constructs a layered representation linking business embeddings to a structured hierarchy of NAICS codes.



Publications & Preprints

Besides Research