ARTFEED — Contemporary Art Intelligence

DotRAG: Training-Free GraphRAG Achieves SOTA on Multi-Hop QA

ai-technology · 2026-05-20

A new framework called DotRAG has been introduced by researchers, which operates without the need for training and redefines retrieval as reasoning through paths. This framework incorporates Division of Thought (DOT) to break down retrieval into specific search areas, tailoring strategies for each query. DotRAG has demonstrated exceptional performance on the MetaQA and UltraDomain benchmarks, consistently outperforming in multi-hop tasks, all while not depending on explicit reasoning chains.

Key facts

  • DotRAG is a training-free GraphRAG framework.
  • It reformulates retrieval as a reasoning process over paths.
  • Division of Thought (DOT) decomposes retrieval into localized search spaces.
  • DotRAG achieves SOTA on MetaQA and UltraDomain.
  • It shows consistent gains on multi-hop tasks.
  • The approach does not use explicit step-by-step reasoning chains.

Entities

Sources