CV

Graduate Research Assistant

Johns Hopkins University, Data Science and AI Institute

Graduate Research Assistant — Scientific Feasibility Assessment Agent

Johns Hopkins University

Advisors: Prof. Benjamin Van Durme, Prof. Anqi Liu

• Architected a modular multi-agent system (using LangGraph/LangChain) that orchestrates iterative loops of literature search, reflection, and summarization to validate materials science claims.
• Identified and formalized the “Contextual Tunneling” failure mode, showing that naive RAG integration can degrade reasoning in scientific domains.
• Led the evaluation strategy and system optimization, achieving top-ranking performance in the DARPA dry run, outperforming five competing research institutions.

Graduate Research Assistant — Enhancing Calibrated Reasoning with Margin Process Reward

Johns Hopkins University

Advisors: Prof. Anqi Liu, Prof. Benjamin Van Durme

• Proposed a calibration-aware RL framework that jointly optimizes reasoning correctness and confidence reliability via margin-based process reward over intermediate reasoning states.
• Designed a lightweight probe-based confidence estimator using Monte Carlo forced-completion as training targets.
• Demonstrated improved calibration across in-domain math and out-of-domain code, science, and logic benchmarks while preserving reasoning accuracy.
• Validated downstream utility of calibrated confidence for conformal risk control and confidence-weighted answer aggregation.

Graduate Research Assistant — Fine-grained Probabilistic Calibration and Reasoning in LLMs

Johns Hopkins University

Advisors: Prof. Benjamin Van Durme, Prof. Anqi Liu

• Developed synthetic data for generous-domain, long-context settings beyond human-annotated datasets.
• Proposed a training framework enabling direct probability modeling with modern LLM backends.
• Constructed an objective, comprehensive evaluation suite from existing datasets and frameworks admitting an intuitive probabilistic interpretation.

Visiting Researcher — Multimodal Clinical Outcome Prediction via LLMs and Mixture-of-Experts

University of North Carolina at Chapel Hill

Advisor: Prof. Huaxiu Yao

• Developed unified noise-resilient encoders to extract information from language descriptions transformed from multi-modal data.
• Employed a sparse Mixture-of-Experts framework to refine and dynamically integrate different modality representations for prediction.
• Achieved state-of-the-art performance predicting clinical trial outcomes on the HINT and CTO datasets.

Visiting Researcher — Predicting Gene Expression from Ultra-Long Sequences using Selective State-Space Models

University of North Carolina at Chapel Hill

Advisor: Prof. Huaxiu Yao

• Constructed a bi-directional Mamba model, a selective state-space sequence model, for gene expression prediction, effectively handling long gene sequences over 100kb.
• Incorporated biological prior information such as peak, exon, and gene locations to enhance model performance.
• Collaborated with a multidisciplinary team to integrate the sequence model into existing bioinformatics pipelines.

Visiting Researcher — Dynamic Backdoor Watermarking for Ownership Protection via Text Embeddings

Hong Kong University of Science and Technology

Advisor: Prof. Minhao Cheng

• Developed a method to protect the ownership of text embeddings by injecting a dynamic backdoor watermark.
• Balanced watermark utility and detectability, even when facing out-of-distribution data.
• Outperformed prior protective approaches in balancing embedding utility and watermark verification.

Software Developer

China Telecom, Shaanxi Branch

PhD

University of Texas at Austin

Computer Science

M.S.

Johns Hopkins University

Computer Science

• Trustworthy Machine Learning; Machine Learning Theory; Intro to Human Language Technology; Self-Supervised Learning

B.Eng.

Xi'an Jiaotong University

Artificial Intelligence (Honor Track)

• Machine Learning; Computer Vision; Natural Language Processing; Modern Control; Digital Signal Processing; Probability Theory and Stochastic Process; Reinforcement Learning

BGA Student (visiting)

University of California, Berkeley

Computer Science

• Optimization Models in Engineering; Designing, Visualizing and Understanding Deep Neural Networks; Data Structures

Visionary Award (Top 30)

2025 LLM Hackathon for Applications in Materials and Chemistry

Excellent

National College Students Innovation Training Program

Provincial Second Prize

National Optoelectronic Design Competition

Provincial Second Prize

Mathematical Contest in Modeling