Xuan (Lily) Yang

I am Xuan (Lily) Yang (杨萱), a PhD student in Computer Science at Duke University, advised by Prof. Jian Pei. I received my bachelor's and master's degrees from Zhejiang University (ZJU), where I collaborated with Prof. Yang Yang. I have interned at Google and TikTok, and visited Stanford and NUS as a research intern.

My research focuses on:

• Data-centric AI — data valuation, selection, and synthesis for trustworthy and efficient models.
• LLM agents / multi-agent systems (MAS) — efficient inference and multi-agent architecture optimization.

📌 Open to collaboration on data-centric AI / MAS projects.
Feel free to reach out — always happy to connect!

Email  /  Google Scholar  /  Linkedin

Batch-of-Thought: Cross-Instance Learning for Enhanced LLM Reasoning    [Code]
Xuan Yang, Furong Jia, Roy Xie, Xiong Xi, Hengwei Bian, Jian Li, Monica Agrawal
We introduce Batch-of-Thought (BoT), a training-free method that processes related queries jointly to identify shared reasoning templates and catch errors through consistency checks, and BoT-Reflection (BoT-R), a multi-agent architecture where a Reflector performs joint evaluation to unlock mutual information unavailable in isolated inference. Experiments across three model families and six benchmarks demonstrate that BoT-R consistently improves accuracy and confidence calibration while reducing inference costs by up to 61%.

D-Shap: Dynamic Shapley Computation
Xuan Yang, Hsi-Wen Chen, Ming-Syan Chen, Jian Pei
We address Shapley-based data valuation in dynamic environments by representing values as a player-by-task matrix and reformulating the problem as matrix maintenance. D-Shap leverages the observation that tasks depend on small subsets of training data and similar tasks yield comparable valuations, enabling targeted updates via structure-aware interpolation. Task updates execute in milliseconds; player updates achieve up to three orders of magnitude efficiency gains while maintaining competitive valuation quality.

TUA-Bench: A Benchmark for Terminal-Use Agents    [Code]
Shoufa Chen, Luyuan Wang, Xuan Yang, Zhiheng Liu, Yuren Cong, Yuanfeng Ji, Feiyan Zhou, Xiaohui Zhang, Fanny Yang, Belinda Zeng
Meta AI, Duke University, Stanford University
TUA-Bench is a general-purpose benchmark evaluating terminal-use agents across 120 tasks spanning five families, including document editing, email management, live-web information seeking, and scientific and engineering workflows. Tasks run in real terminals with deterministic setup scripts and execution-based scoring.

Local Shapley: Efficient Data Valuation for Model Training    [Code]
Xuan Yang, Hsi-Wen Chen, Ming-Syan Chen, Jian Pei
We propose Local Shapley, which formalizes model-induced locality through support sets to focus Shapley computation on influential training points rather than exhaustive coalition enumeration. We design and prove that LSMR achieves the optimal number of model training runs by training each support exactly once; LSMR-A extends this with an unbiased Monte Carlo estimator for larger supports. Experiments across multiple model families demonstrate substantial retraining reductions and speedups while preserving high valuation fidelity.

Unfolding and Modeling the Recovery Process after COVID Lockdowns    [News]
Xuan Yang, Yang Yang, Chenhao Tan, Yinghe Lin, Zhengzhe Fu, Fei Wu, Yueting Zhuang
Nature Scientific Reports, 2022 (Front-Page Feature, Zhejiang Daily)
We present a graph-learning–based computational framework leveraging electricity consumption data to analyze post-lockdown recovery dynamics. Our approach quantifies sector-specific impacts, evaluates the effectiveness of government recovery policies, and models inter-sector dependencies to inform more effective strategies for holistic economic revitalization.

Who's Next: Rising Star Prediction via Diffusion of User Interest in Social Networks   
Xuan Yang, Yang Yang, Jintao Su, Yifei Sun, Shen Fan, Zhongyao Wang
IEEE Transactions on Knowledge and Data Engineering, 2022
We propose RiseNet, a novel recommendation framework designed to identify potential “Rising Star” items and mitigate unfairness in recommendation systems. RiseNet models the dynamic diffusion of user interests alongside temporal item features, using a coupled mechanism to capture their interactions and a multi-task GNN-based framework to quantify user interest. Experiments on real-world Taobao data demonstrate its effectiveness in predicting emerging popular items.

DropMessage: Unifying Random Dropping for Graph Neural Networks   
Taoran Fang, Zhiqing Xiao, Chunping Wang, Jiarong Xu, Xuan Yang, Yang Yang
AAAI, 2023 (Distinguished Paper Award)
We present a unified framework that generalizes existing random dropping techniques by applying dropping operations to the message matrix in Graph Neural Networks (GNNs). Building on this, we propose DropMessage, a versatile method applicable to any message-passing GNN. Theoretically, DropMessage improves training stability by reducing sample variance and enhances information diversity from an information-theoretic perspective.

Google    Kirkland, WA
Summer Intern, Google Cloud team
May 2026 -- Present

TikTok    Bellevue, WA
Research Intern, Risk Control team
May 2025 -- Apr 2026

Alibaba Group    Hangzhou, China
Research Intern, Data Assets and Algorithm team
Oct 2020 -- Dec 2021

Stanford University    Palo Alto, CA
Research Assistant, Center for Magnetic Nanotechnology
Jan 2019 -- Mar 2019

National University of Singapore    Singapore
Research Assistant, Big Brain, BIGHEART
Jun 2018 -- Aug 2018