Biography
Hi! 👋🏻 I am Xunyi Jiang, a 1st year PhD student at NYU advised by Hongyi Wen. Previously, I was fortunate to be advised by Prof. Julian Mcauley at UCSD during my master’s studies and Prof. Yifang Ma during my undergraduate at SUSTech.
I was always facinated by understanding complex systems and their behaviors. My recent research focus is on LLM understanding and evaluation, Human-AI collaboration, AI for creativity🎻, and memory- augmented LLMs.
I love music! This is one of my passion apart from coding and math. I play both cello and piano, and I am a member of GG Ochestra. I used to be a principle cellist of the SUSTech Symphonic Orchestra. I also produce music and drawing in my free time.
Feel free to reach out to me via email for any collaboration or just to say hi! 😊
- LLM Evaluation and Understanding 🤖
- Human-AI Collaboration 🧑💻
- AI for Creativity and Music 🎹
Ph.D. in Computer Science, 2026-
New York University
M.S. in Computer Science, 2024-2026
University of California, San Diego
B.S. in Data Science, 2020-2024
Southern University of Science and Technology
High School, 2017-2020
No.1 Affiliated Middle School of Central China Normal University
News🔥
Publications
Research
Working on:
LLM Evaluation and Understanding
Evaluating LLM training dynamics.Human-AI Collaboration
Exploring methods to enhance collaboration between humans and AI systems in Deep Research, focusing on improving user experience and task efficiency.
Working on:
Trustworthy LLMs
Exploring methods to mitigate hallucinations in large language models, ensuring the reliability and accuracy of generated content.Controllable Symbolic Music Generation
Developing Steering techniques to guide the music generation process, allowing for confusion and continuous control over the output style.Music Editing Evaluation
Assessing the effectiveness of music editing tools and techniques in enhancing the creative process.
CLUE-ReDial: Leveraging Large Language Models for Generating Comprehensive Dataset with URLs and Explanation
View detailed responsibilities
- Designed and implemented a multi-agent framework leveraging RAG and LLMs, achieving a 4× efficiency boost in generating interpretable recommendation explanations. Expanded the ReDial dataset to 40k high-quality utterances for fine-tuning conversational LLMs, enhancing domain-specific accuracy in movie recommendations.
- Designed a multi-agent framework, leveraging an early-stop module in the self-refine process to generate comprehensive and interpretable recommendation explanations, boosting transparency in recommendations.
- Applied web crawling techniques to create a movie dataset containing 25k entries, using Python, PyTorch, LaTeX, and the Singapore National High-Performance Computing (HPC) platform.
Worked on the OpenAlex and ORCID databases to analyze the relationship between university rankings and the mobility of researchers.
View detailed responsibilities
- Extracted education and hiring networks from research paper publications.
- Ranked Chinese universities based on extracted database information.
- Implemented a Multi-Variate Regression (MVR) model for ranking.
- Applied Markov Chain Monte Carlo (MCMC) and burn-in techniques to ensure the stability of convergence.
- Visualized researcher mobility and constructed a null model to compare the hierarchical structures of research institutions.
Skills
Accomplishments
Recent Posts
How to merge external knowledge into LLMs
Xunyi Jiang Southern University of Science and Technology xunyijiang001@gmail.com
This report explores the evolving landscape of Conversational Recommender Systems (CRS) in the context of Large Language Models (LLMs). It delves into the integration of knowledge graphs and advanced computational techniques to enhance the capabilities of CRS. The report categorizes current CRS methods into four main classes: knowledge-based, goal-driven, fine-tuning, and agent-based approaches, providing an in-depth analysis of each. Special attention is given to how these systems can optimize conversations and recommendations by harnessing the power of LLMs and external knowledge sources. The innovative use of agent-based methods and the challenges of fine-tuning LLMs for CRS are also discussed, highlighting the importance of efficient retrieval methods and the integration of external knowledge for improved system performance.
Popular Topics
Contact
- xunyijiang001@gmail.com
- 60 5th Ave, New York, NY 10011, New York, New York
- Courant Institute of Mathematical Sciences
- DM Me
- Zoom Me