RSF-GLLM bridges semantic gap in multi-hop KG QA via recurrent soft-flow

The authors propose RSF-GLLM, a framework that decouples differentiable graph reasoning from answer generation to address the lack of lexical overlap in multi-hop Knowledge Graph Question Answering. The Recurrent Soft-Flow (RSF) module uses a GRU-guided query updater and dynamic gating to traverse semantically dissimilar bridge nodes via structural cues.

• RSF propagates continuous relevance scores to bridge the semantic gap where traditional pipelines break differentiability.
• Flow sparsity regularization theoretically guarantees convergence from soft probabilities to discrete reasoning paths.
• Extracted paths are textualized to fine-tune a Large Language Model, grounding generation in factual topology.
• Experiments on WebQSP and CWQ show competitive performance with superior inference efficiency compared to computationally expensive LLM-based approaches.

The framework ensures generation is grounded in factual topology while achieving competitive performance with superior inference efficiency.