SPLIT-RAG at EMNLP 2026: multi-agent graph partitioning cuts agentic RAG overhead

flosalim · x · 2026-09-21

The author presents work on robust and efficient agentic retrieval accepted at EMNLP 2026, led by student Ruiyi Yang (first author on 3 of 6 accepted papers). SPLIT-RAG (Main) tackles the search overhead, irrelevant evidence and conflicting facts that arise when complex multi-hop queries retrieve over large knowledge graphs. The multi-agent framework includes: question-driven graph partitioning, which organizes a large KG into compact, reusable partitions aligned with recurring reasoning patterns; selective multi-agent retrieval, where each query activates only a few partition-specialized agents retrieving in parallel; and compatibility-based aggregation of retrieved evidence.

Original post →

More from coding & agent

coding & agent channel →