Research
RSF-GLLM Preserves Differentiability in Multi-Hop Knowledge-Graph QA
RSF-GLLM replaces brittle retrieve-then-read pipelines for multi-hop knowledge-graph QA with a Recurrent Soft-Flow mechanism plus decoupled LLM generation, aiming to keep the pipeline differentiable across reasoning hops. Relevant to builders doing GraphRAG/KG-QA where multi-hop reasoning currently breaks at the retrieval boundary. The soft-flow evidence routing is the reusable component.
Source
↳ Follow the thread