P Moth-Retrieval: Graph-Free Multi-Hop Retrieval via Query-Time Orchestration (Beating Graph-Based Systems on HotpotQA) [P]

Summary

We just open-sourced MOTHRAG, a multi-hop RAG framework that skips the knowledge graph entirely. We kept hitting the same wall building multi-hop RAG: the systems with the best accuracy (GraphRAG, HippoRAG, RAPTOR) all lean on a knowledge graph built offline, and that’s great numbers, until the moment your data changes! Every single update means re-running a heavy LLM indexing pass to rebuild the graph. If your corpus updates daily (prices, internal filings, support tickets, news), you're pa

SpinGraph analysis pending — check back after processing.

• A system-level approach to prompt injection: separating instruction and data channels in LLM agents [P]
• How to describe a model that has higher accuracy with fewer #param and FLOPs? [D]
• ICML qr code visible [D]
• A map of the latest 11 million papers split by semantic similarity and time slices [P]
• [D] Simple Questions Thread
• Making Optimization Work When Labels Are Scarce [R]