Paper Reading Night - Week 27 - TBD

​This week's paper

​Tentative topic https://arxiv.org/html/2405.20139

​What are the group goals? Stay on top of AI research and improve our understanding of AI fundamentals, retrieval-augmented generation, graph neural networks, and LLM reasoning.

​Who is welcome? Everyone! Try to spend some time with the paper beforehand and come prepared with questions or ideas you'd like to discuss—but it's also okay to simply show up and learn with the group.

​About the paper

​Retrieval-augmented generation usually retrieves passages of text, but many questions require reasoning across relationships spread throughout a knowledge graph. GNN-RAG combines the complementary strengths of graph neural networks and large language models: a GNN searches a dense knowledge-graph subgraph for likely answers and useful multi-hop reasoning paths, then verbalizes those paths as context for an LLM.

​The approach achieves state-of-the-art results on WebQSP and Complex WebQuestions. It is especially effective on multi-hop and multi-entity questions, outperforming competing approaches by 8.9–15.5 percentage points in answer F1. Remarkably, the system can match or outperform GPT-4 using a tuned 7B language model—showing how better retrieval and structured reasoning can sometimes matter more than model scale.

​Read the full paper on arXiv: https://arxiv.org/html/2405.20139