Reproduce “LatentRAG: Latent Reasoning and Retrieval for Efficient Agentic RAG” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yijia Zheng, Marcel Worring

LatentRAG represents intermediate thoughts and search queries as continuous hidden states for multi-hop question answering, rather than generating them as text. Its latent queries retrieve documents through a retriever trained to match reference distributions. Across seven QA datasets, the Search-R1-trained Qwen2.5-7B variant averages 43.46% exact match and 593 ms latency, versus 42.47% and 5,372 ms for Search-R1, an 89.0% latency reduction; retrieval itself is not shown to be faster.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.