cv

Master of Science in Computer Science

University of California, San Diego, California, USA

• GPA: 4.00/4.00

Bachelor of Science in Computer Science and Technology

Central South University, Hunan, China

• GPA: 90.97/100 (3.88/4.00)
• Awarded the School of Computer Science and Engineering Annual Scholarship (2018 - 2021)
• Awarded 2022 Outstanding Graduates of Central South University (Jun. 2022)

PromptAgent - Automatic LLM Prompt Optimization Framework

University of California, San Diego, California, USA

• Motivation: Existing prompt optimization methods tend to overlook the depth of domain knowledge and struggle to efficiently explore the vast space of expert-level prompts using simple sampling or searching methods. PromptAgent formulates prompt optimization as a strategic planning problem and employs a principled planning algorithm to navigate the expert-level prompt space efficiently and effectively.
• Methodology: Introduced a novel automatic prompt optmization agent that integrates Monte Carlo Tree Search (MCTS), a principle planning algorithm, to explore the prompt space. Additionally, by employing a trial-and-error approach, the domain knowledge in the data is incorporated into the optimzed prompt during optimizing.
• Results: PromptAgent has been rigorously experimented on various tasks, covering BIG-Bench Hard, domain-specific, and general NLP tasks, showing it outperforms strong human prompting methods, such as Chain-of-Thought, as well as recent prompt optimization baselines. Analyses highlight its ability to craft expert-level, detailed, and domain-insightful prompts with superior performance and efficiency than other baselines. Moreover, the optimized prompts demonstrate better generalizability across models compared to those written by humans or other baselines.
• Output paper: PromptAgent - Strategic Planning with Language Models Enables Expert-level Prompt Optimization

LLM Reasoners Library

University of California, San Diego, California, USA

• Motivation: LLM Reasoners is a library designed to enhance the reasoning capabilities of Large Language Models (LLMs) using advanced algorithms. It treats multi-step reasoning as a planning task, searching for the optimal reasoning chain. This approach aims to strike the right balance between exploration and exploitation, guided by concepts like the 'World Model' and 'Reward'.
• Methodology: The library is based on Reasoning via Planning. It offers the most up-to-date search algorithms for reasoning with LLMs, such as RAP-MCTS, Tree-of-Thoughts, Guided Decoding, and more. Additionally, the framework is compatible with various LLM platforms, including Huggingface transformers, OpenAI API, and others. I am responsible for 1. Managed experiments on Math datasets using various reasoning methods, including Least-to-Most Prompting, CoT, and RAP-MCTS. 2. Developed Tree-of-Thought searching algorithm.
• Results: This library integrates cutting-edge LLM reasoning algorithms, the latest LLM models, and intuitive visualization. It also offers accessible interfaces for all incorporated algorithms and models. Notably, among the recent LLM reasoning algorithms, our RAP-MCTS algorithm shows superior performance over other searching or prompting methods on Math, Logical, and Embodied tasks.
• Github Library: LLM Reasoners Github Stars: 521
• Paper: New Evaluation, Library, and Analysis of Step-by-Step Reasoning with Large Language Models

Refining Diffusion Model Loss with End-to-End Information

University of California, San Diego, California, USA

• Motivation: Diffusion models can generate photorealistic images, but they often make many semantic mistakes, such as twisted fingers and ears. This is because Diffusion models are trained with per-pixel loss (MSE) in an auto-regressive manner, which lacks end-to-end conceptual information. Incorporating end-to-end information into the training loss of diffusion models is essential to correct the conceptual errors and enhance performance of diffusion models.
• Methodology: A Generative Adversarial Model (GAN) is concurrently trained with the Diffusion Model, utilizing end-to-end synthesized data. The discriminator of GAN injects conceptual guidance into Diffusion models' training process. This approach allows the original per-pixel loss to be augmented with comprehensive set information.
• Results: Experiments show that the unconditional transformer-based Diffusion model, DiT, can generate images of higher fidelity (measured by FID) when trained with end-to-end conceptual guidance on CelebA-HQ Dataset.

Interpretable Object Detection via Deep Learning

Central South University, Hunan, China

• Motivation: CNNs are widely used as black boxes, while the interpretability of their features and feature importance is less explored. This project aimed to improve the interpretability of CNN by visualization and provide guidance for feature transferring of high-value features in CNN.
• Methodology: Proposed the Average Image Analysis method to evaluate the network's low-frequency information. It measures the input's influence on neurons by calculating the cosine distance between neuron's guided-backpropagation visualizations and the dataset's average image.
• Results: This approach improves the clarity of gradient-based CNN visualizations, deepens insights into semantic information at different network depths, and provides guidance for subsequent feature transfer. Experiments reveal that pruning 40% of high-importance features dramatically reduces performance, while eliminating 40% of low-importance features has negligible impact on the model's performance.
• Software Copyright: A Visual Analysis of Internal Feature Importance and Feature Transfer Method in Convolutional Neural Networks, 2021-06-25
• Graduation Thesis: The Research on The Interpretability Method of Deep Neural Network Based on Average Image, 2022-06-01

Automatic Medical Triage System Based on Natural Language Processing

Central South University, Hunan, China

• In line with solving practical medical problems, I assisted with researching medical triage issues in China and with my classmates, built an automatic medical triage system based on medical question answering data using Bidirectional Encoder Representation from Transformers
• Our system performs well when testing with 95% top-2 accuracy on 5 major disease data sets and 78.3% top-2 accuracy on 20 disease data sets, which is significantly higher than Sequence-to-Sequence models built by others
• Output paper: “Reduce the medical burden: An automatic medical triage system using text classification BERT based on Transformer structure”, ICBASE 2021

Detecting Malicious Webshells by Constructing Webshell Family Clusters

Central South University, Hunan, China

• Webshells are malicious scripts in web servers. This project aims to detect malicious webshells by constructing webshell families according to the clusters of their program call sequences. The main method is composed of reducing the dimensionality of webshell code with Auto-encoders, clustering low-dimensional webshells and interactively constructing webshell families using Visualization techniques.
• I researched on text classification CNN, implemented and tested different CNN-based Auto-encoders and applied different clustering algorithms. I also merged dimensionality reduction and clustering into one model using Contrastive Dimensionality Reduction model.
• Software Copyright: “Interactive Family-building Software of Malicious Webshell Files V1.0”, 2022-06-27

Visualization of minority structures in graphs

Central South University, Hunan, China

• Minority structures in a graph, like high degree nodes and bridges, are important for graph analysis and comprehension.
• I proposed two kinds of minority graph structures called super pivot and huge star and one algorithm to detect these structures by identifying triangle structures using depth-first-search. Experiments showed that this algorithm worked well on small- and middle-scale graphs with high accuracy, recall value and time efficiency.
• VINCI 2020 poster, “A fast method for detecting minority structures in a graph.”. In the proposal, We proposed four typical minority structures in graphs and two algorithms to detect these structures fast and efficiently.

• School of Computer Science and Engineering Annual Scholarship

• School of Computer Science and Engineering Annual Scholarship

• School of Computer Science and Engineering Annual Scholarship
• Guangyun Technology Scholarship

• 2022 Outstanding Graduates of Central South University