CV

Ph.D.

University of Virginia

Computer Engineering

• {“Advisor”=>”Prof. Jundong Li.”}

B.E.

University of Science and Technology of China

Computer Science

Research Advisor II

Meta

Topic: LLM for Recommendation. Adviser: Dr. Cassie Dong.

• Built a large-scale transformer-based model for effective and efficient ads ranking.

Machine Learning Research Intern

Nokia America

Topic: Generative Recommendation. Adviser: Dr. Liang Wu.

• Studied semantic ID (SID) based generative recommendation with large language models, focusing on reasoning-aware generation and efficient inference.
• Constructed explicit Chain-of-Thought (CoT) reasoning traces for each meta token in SID generation and trained a teacher model with a two-stage SFT-then-RL pipeline.
• Introduced auxiliary soft-thinking tokens for approximately 1B-parameter student models and distilled the teacher’s CoT into continuous latent embeddings.
• Deployed only the student model at inference time, achieving comparable ranking quality to explicit CoT-based generation with reduced latency and generation cost.

Machine Learning Research Intern

Snap Inc.

Topic: Agentic Recommendation Systems. Adviser: Dr. Clark Ju.

• Developed a personalized LLM-based agent for multi-channel item recall by extending Hugging Face’s TRL framework.
• Designed and implemented a soft-token mechanism to structurally generate routing weights over recall channels.
• Modified the vLLM inference engine to support high-throughput soft-token decoding and agent-style online interaction.
• Adapted Group Relative Policy Optimization (GRPO) to optimize adaptive routing over multiple recall models.

Research Assistant

Snap Inc.

Topic: Cold-Start Item Recommendation. Adviser: Dr. Neil Shah.

• Developed an adaptive cold-start item recommendation system for rapidly emerging items with sparse or zero interaction histories.
• Proposed MI4Rec, a PLM-based cold-start recommender that infers item representations via meta-item embedding composition.
• Evaluated on Amazon-Beauty, Toys, Sports, and Yelp under cold-start, warm-start, and general settings, achieving up to 15.71% relative improvement in Recall.

Machine Learning Engineer Intern

Zerozero Robotics

Topic: Trajectory Planning. Adviser: Hao Zhang.

• Built an outdoor trajectory planning system for physical quadrotors and constructed realistic evaluation environments in Unity.
• Reproduced an imitation learning framework and incorporated data aggregation to improve obstacle avoidance.
• Achieved an 18.2% relative reduction in navigation failure rate in outdoor environments.

Research Assistant

University of Virginia

Topic: Retrieval-Augmented Generation (RAG). Adviser: Prof. Jundong Li.

• Proposed CoRAG, a hybrid RAG framework that dynamically selects between global textual retrieval and graph-based relational retrieval.
• Designed a Cooperative-Retrievers (CoR) mixture-of-retrievers module with hierarchical gating to adaptively fuse heterogeneous retrieval signals.
• Demonstrated strong performance on STaRK semi-structured KGQA benchmarks, achieving up to 11.4% relative Hit@1 improvement over prior RAG baselines.

Research Assistant

University of Texas at Austin

Topic: Indoor Navigation. Adviser: Dr. Wuyang Chen.

• Developed an efficient and transferable robot navigation method for sim-to-real deployment.
• Implemented a PPO-based navigation pipeline on Meta’s Habitat platform.
• Proposed a contrastive learning component for visual representation learning with positive and negative sampling.