cv

Master of Science (Thesis Track)

Bachelor of Science (Honor Class)

Trustworthy AI Code Generation (In Progress)

• Proposed a Tree of Thought-based framework using Large Language Model (LLM) APIs, designed to enhance code generation quality through the preservation of both semantic consistency and diverse expression.
• Demonstrated the stubbornness of current code-LLMs through an extensive research investigation on model architectures, finetuning strategies and prompting strategies.
• Collaborated with Prof. Junfeng Yang, Baishakhi Ray, Zhou Yu, Kexin Pei, and Chengzhi Mao.

Symmetry-Presering Program Representation for Learning Code Semantics

• Proposed a foundational framework using semantics-preserving code symmetries, ensuring provable generalization to new samples resulting from their compositions.
• Presented a novel mechanism within LLMs architecture, featuring a unique variant of self-attention that is equivariant to program symmetries, identified through graph automorphism.
• Conducted practical experiments with multiple implementation variants to translate theoretical concepts into applicable solutions.
• Demonstrated effectiveness in generalizing invariance to various semantics-preserving transformations, surpassing state-of-the-art code LLMs in various program analysis tasks.
• Collaborated with Prof. Junfeng Yang and Kexin Pei.

Learning-based Automated Test Generation

• Proposed a new dimension that leverages LLMs to generate assertions for incomplete test methods in an infilling style. Enhanced alignment with LLMs' pre-training objectives and utilized bidirectional context for smoother integration.
• Fine-tuned LLMs using various masking strategies at different granularities to ensure precise hidden tokens reconstruction.
• Conducted extensive evaluations to estimate the differential impact of various masking strategies, and the performance across distinct LLMs, alongside an analysis of operational costs.
• Collaborated with Prof. Lingming Zhang.

Beyond Entities: A Large-Scale Multi-Modal Knowledge Graph with Triplet Fact Grounding

• Proposed the first-ever framework for grounding triplet facts in images, introducing a new dimension of visual semantics of relations to multimodal knowledge graphs (MMKGs).
• Constructed ImgFact, a groundbreaking large-scale MMKG, employing a retrieval-based approach to assemble an extensive dataset comprising 247K triplets and 3M images.
• A series of meticulous manual and automated evaluation processes has been conducted, demonstrating the dataset's exceptional quality and reliability.
• Leveraged the ImgFact dataset to significantly enhance model performance across various real-world applications. The model optimized by ImgFact achieves an impressive 8.38% and 9.87% improvement in F1 score over the solutions enhanced by an existing MMKG and VisualChatGPT, respectively, on relation classification.
• Collaborated with Prof. Yanghua Xiao.

Vanilla CPU @ 4th National CPU Design Contest

• Designed and implemented an experimental dual-issued MIPS CPU that supports around 80 MIPS instructions and all required CP0 registers in MIPSr1, as well as various exception handling mechanisms.
• Achieved over 30% speedup by implementing Harvard architecture memory hierarchy and refactoring the official AXI bridge to support and optimize burst transfers.
• Implemented TLB memory management and capable of running whole PMON system and uCore debug mode.
• Ranked 6th and Second Prize Awardee, collaborated with Prof. Liang Zhang and Chen chen.

Deep SEM

• Researched and studied the Neural Architecture Search (NAS) and a social science research approach named Structure Equation Model (SEM).
• Designed and implemented an experimental framework to apply NAS and reinforcement learning to speed up the traditional SEM process from days to hours.
• Used PyQt5 to build a frontend that enabled users to import questionnaire data and construct initial relationship.

• Silver Medal @ International Collegiate Programming Contest East Continent Final

• Gold Medal @ China Collegiate Programming Contest Changchun Site
• Gold Medal @ China Collegiate Programming Contest WFinal
• Rank 6th @ 4th National Student Computer System Capability Challenge
• Google APAC Women Techmakers Scholarship 2020

• Third Prize @ 35th National Olympiad in Informatics, China

• Analyzed user behaviors of the GPay payment transaction, proposed and implemented a new QR code profile widget, and improved the payment transaction by at least 10% in the production environment.
• Refactored the existing EMV QR code parser pipeline to ease the logic of integrating a new EMV QR. Implemented the QR code parsing pipeline that launched in Brazil.
• Constructed a new object detection dataset for QR code and spot code and applied the resampling mechanism to improve the code detection neural network model by 5% on accuracy.

• Analyzed the time consumption of each part of the derivatives statement generating system and implemented a scheduler to eliminate the bottleneck and improve the efficiency of the derivatives statement generating system by nearly 37%.
• Researched the basic principles of transaction management in SpringBoot, gave a small lecture to the group, and applied it to the derivatives statement management system.