School of Engineering Department of Computer Science and Engineering 143 Finding Bugs in Compilers with/without AI Techniques Supervisor: TIAN Yongqiang / CSE Student: ZHONG Yingqi / COMP Course: UROP 1100, Fall Building upon the prior work of replicating CCMD mutation operators and under the guidance of the project supervisor, the student focused on designing a new mutation operator aimed at improving the generation of debug information. This effort was based on a literature review and a white-box analysis of debug information and prior mutation operators. The analysis discussed the constraints and rationales of the existing mutation operators and the methodology used to design the new operator. The work also included a high-level analysis discussing potential future directions for CCMD-related studies. Efficient Queries over Database Supervisor: WONG Raymond Chi Wing / CSE Student: CHEN Yifu / COSC Course: UROP 1000, Summer In real-world spatial networks, road travel times often follow various distributions that can change dynamically due to factors such as individual driving speed, time of day, and weather conditions. These fluctuations can impact travel efficiency, prompting users to prefer faster routes or those with more consistent arrival times. Therefore, a path that balances speed and time stability is essential. Additionally, traffic managers need to update travel time distributions in response to significant disruptions, such as traffic congestion, accidents, or emergencies, which can cause abrupt changes in travel times. This report focuses on designing an algorithm that calculates the reliable shortest path based on user preferences for speed or stability and efficiently processes updates to travel time distributions. Efficient Queries over Database Supervisor: WONG Raymond Chi Wing / CSE Student: DONG Yunao / DSCT Course: UROP 2100, Fall UROP 3100, Spring UROP 1000, Summer In this project, I collaborated with PhD student TAN, Weile and Professor Raymond Wong Chi Wing to design and implement a multimodal recommendation framework based on pre-trained large language models. My work done this semester is as follows. First, I conducted an in-depth analysis of cutting-edge vision-language architectures, including DeepSeek-VL and Qwen-VL, dissecting their vision encoder mechanisms to guide the design of a multimodal encoding interface between recommendation systems and LLMs. Then, I validated LLaRa (Large Language-Recommendation Assistant), a parameter-efficient fine-tuning framework optimized for adapting pre-trained LLMs to recommendation scenarios, which was successfully deployed on server infrastructure. These foundational efforts—spanning requirement analysis, architecture benchmarking, and technical validation—establish a pipeline for building an end-to-end system that combines multimodal feature extraction with LLM-driven reasoning.
RkJQdWJsaXNoZXIy NDk5Njg=