TreeHop: Generate and Filter Next Query Embeddings Efficiently for Multi-hop Question Answering
Zhonghao Li, Kunpeng Zhang, Jinghuai Ou, Shuliang Liu, Xuming Hu
2025-04-30
Summary
This paper talks about TreeHop, a new way for AI to answer complicated questions that need information from several different places, while using less computer power.
What's the problem?
When AI tries to answer multi-step questions, it usually has to keep asking itself new questions and checking lots of sources, which takes a lot of time and computer resources.
What's the solution?
The researchers designed TreeHop to work at the level of embeddings, which are like smart summaries of information, so the AI can quickly figure out what to look up next without having to constantly go back and forth with big language models.
Why it matters?
This matters because it makes AI much faster and cheaper to use for answering tough questions, which is helpful for things like research, studying, or any situation where you need to connect lots of facts together.
Abstract
TreeHop, an embedding-level framework for multi-hop question answering, significantly reduces computational costs compared to traditional retrieval-augmented generation systems by eliminating the need for iterative LLM-based query refinement.