Sambaran Bandyopadhyay, Ananth Muppidi
A differentiable graph reasoning framework for multi-hop knowledge graph QA that bridges the semantic gap by decoupling recurrent soft-flow from LLM generation.
Traditional retrieve-then-read pipelines for multi-hop knowledge graph QA break differentiability, preventing the retriever from learning to bridge the semantic gap where intermediate nodes lack lexical overlap with the query.
The Recurrent Soft-Flow (RSF) module uses a GRU-guided query updater to propagate continuous relevance scores and a dynamic gating mechanism to traverse semantically dissimilar bridge nodes via structural cues. Flow sparsity regularization theoretically guarantees convergence from soft probabilities to discrete reasoning paths, which are extracted and textualized to fine-tune an LLM.
Experiments on WebQSP and CWQ show RSF-GLLM achieves competitive performance with superior inference efficiency compared to computationally expensive LLM-based approaches. The work introduces a novel architecture that effectively decouples differentiable graph reasoning from answer generation to address the semantic gap problem.