Zheyi Xue

• Designed and implemented a systematic benchmark comparing pure graph retrieval, agentic dense retrieval, and agentic graph retrieval paradigms for GraphRAG systems
• Reproduced 5 GraphRAG methods end-to-end (RAPTOR, GraphRAG, HypergraphRAG, LinearRAG and HippoRAG2) covering: 1) Corpus selection and preprocessing; 2) Graph schema design and construction; 3) Retrieval strategies and LLM inference pipelines
• Evaluated methods on both single hop and multi-hop QA benchmarks (Natural Questions and HotpotQA) across multiple LLM backbones (Llama, Qwen, Mistral, Gemma), enabling controlled cross-model analysis
• Extended experiments to additional datasets (PopQA, TriviaQA and MuSiQue) to study generalization and retrieval robustness
• Conducted fine-grained error analysis to understand performance differences attributable to graph structure and retrieval mechanisms

Keywords:Graph RAG, Knowledge Graph, RAG, Multi-hop Reasoning, Becnmark, Graph Construction

Conducted an in-depth study of graph structure contribution in GraphRAG systems inspired by GraphICL: Unlocking Graph Learning Potential in LLMs through Structured Prompt Design.

• Reproduced LLaGA (LLaGA: Large Language and Graph Assistant) by training from scratch on 4 datasets: Arxiv, Amazon-Products, PubMed and Cora
• Designed controlled ablation experiments by replacing k-hop neighbor subgraphs and Laplacian-based structural signals with semantic similarity-based node selection
• Retrained models under identical settings to isolate the impact of explicit graph structure vs semantic-only context
• Demonstrated that graph structural information remains critical for LLaGA performance across datasets
• Authored internal technical reports analyzing workflows, assumptions, and empirical results of 10+ GraphRAG-related papers

Keywords: Knowledge Graph, Graph Language Models, Ablation Study, Graph RAG

• Reproduced VideoRAG (VideoRAG: Retrieval-Augmented Generation with Extreme Long-Context Videos) for extremely long video (1 h+) comprehension, focusing on graph-enhanced retrieval over multimodal content
• Conducted quantitative comparisons against baselines including NotebookLM and VideoAgent under offline HPC constraints
• Adapted the VideoRAG pipeline to use Gemma-2-9B instead of GPT-4o-mini due to restricted online API access, enabling fully local inference
• Implemented lightweight architectural modifications to improve retrieval and reasoning performance on selected evaluation subsets
• Analyzed bias and variance introduced by limited data regimes, documenting limitations and trade-offs in experimental design

Keywords: Multimodal RAG, Video Understanding, Graph RAG, Long-Context LLMs