Current Focus

Focus on post-training, but I’d rather say I’m focusing on teaching model how to do things and building its personality to behave properly.

Currently we are at the stage of using agent as driving Tesla, but the future should be we’re using agent as we’re taking Waymo. The gap is between: good tool for model’s execution, which is harness; good eval for model’s capability, which needs the domain knowledge; good boundary for model’s capability, which is guardrails,

Education and Research

AI Master’s student at Carnegie Mellon University.

Major: Master of Science in Artificial Intelligence and Innovation

I am working with Prof. Graham Neubig and Prof. Sherry Tongshuang Wu for LLM and Agent research.

🧠 Agent Workflow Memory Management: We are developing a hierarchical memory framework for LLM-based agents, optimizing long-horizon task planning by structuring and managing workflows in memory to reduce context length and computational overhead.
✈️ Travel Planner Agent: Developed a multi-agent travel planner that dynamically updates itineraries by tracking user location and integrating real-time factors such as traffic, and weather, ensuring a seamless and adaptive travel experience. Exploring methods to leverage LLMs for inferring user preferences from implicit behaviors and incorporating external factors, enabling their integration as constraints in agent-based planning for personalized decision-making.
📊 Synthetic data for sft: we designed a data processing pipeline and a RAG method to use Large Language Model 🦙 to address task-specific data shortages and extend these solutions to improve self-alignment.

B.S. at Nankai University.

Major: Software Engineering

I am fortunate to work as a research assistant in the AIOps lab under the guidance of Prof. Shenglin Zhang. In addition, I work as a research assistant in the Human-Computer Communication Laboratory at the Chinese University of Hong Kong under the guidance of Prof. Meiling Meng. I focus on research in Speech, NLP, and Human-Computer Interaction.

Publication

SELF-GUIDE: Better Task-Specific Instruction Following via Self-Synthetic Finetuning (https://arxiv.org/abs/2407.12874)
Training Task Experts through Retrieval Based Distillation (https://arxiv.org/abs/2407.05463)