RSF-GLLM: Bridging the Semantic Gap in Multi-Hop Knowledge Graph QA via Recurrent Soft-Flow and Decoupled LLM Generation

2026-07-08 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, extended

Summary

RSF-GLLM is a novel framework designed to address the semantic gap in multi-hop Question Answering over Knowledge Graphs (KGQA) by decoupling differentiable graph reasoning from Large Language Model (LLM) answer generation. Its Recurrent Soft-Flow (RSF) module employs a GRU-guided query updater and a dynamic gating mechanism to propagate continuous relevance scores, traversing semantically dissimilar bridge nodes via structural cues. Flow sparsity regularization guarantees convergence to discrete reasoning paths. These paths are then extracted and textualized to fine-tune an LLM, ensuring factual grounding. Experiments on WebQSP and CWQ benchmarks show RSF-GLLM achieves 90.45% Hit@1 on WebQSP and competitive performance on CWQ, with a 21x inference speedup over computationally expensive LLM-based agentic approaches, requiring only a single NVIDIA A100 GPU.

Key takeaway

For AI Scientists and Machine Learning Engineers developing multi-hop Knowledge Graph Question Answering systems, RSF-GLLM offers a compelling alternative to computationally expensive agentic LLM approaches. You can achieve competitive accuracy and significantly reduce inference latency by adopting its decoupled reasoning-generation architecture. Consider integrating its Recurrent Soft-Flow module to ensure factual grounding and mitigate LLM hallucinations, especially when dealing with queries requiring traversal through semantically dissimilar intermediate nodes.

Key insights

RSF-GLLM decouples graph reasoning from LLM generation to bridge semantic gaps and ensure factual grounding in multi-hop KGQA.

Principles

• Decouple graph reasoning from LLM generation for efficiency.
• Use dynamic gating to adaptively balance structural and semantic cues.
• Apply flow sparsity regularization to ensure discrete, interpretable reasoning paths.

Method

The Recurrent Soft-Flow (RSF) module uses a GRU-guided query updater and dynamic gating to propagate continuous relevance scores, regularized for sparsity, then extracts paths to fine-tune an LLM.

In practice

• Achieve 21x inference speedup over agentic LLM baselines.
• Reduce LLM hallucination via grounded path generation.
• Run multi-hop KGQA on a single NVIDIA A100 GPU.

Topics

• Knowledge Graph Question Answering
• Multi-hop Reasoning
• Large Language Models
• Graph Neural Networks
• Semantic Gap
• Retrieval-Augmented Generation

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Leveraging LLM-GNN Integration for Open-World Question Answering over Knowledge Graphs — Integrating GNN-predicted candidates and KG context into LLM prompts significantly improves open-world question answering over incomplete knowledge graphs.
• GRISP: Guided Recurrent IRI Selection over SPARQL Skeletons — GRISP uses a fine-tuned SLM for efficient, accurate knowledge graph question answering via iterative SPARQL skeleton generation and IRI selection.
• LogosKG: Hardware-Optimized Scalable and Interpretable Knowledge Graph Retrieval — LogosKG offers scalable, interpretable k-hop retrieval for large knowledge graphs via hardware-optimized symbolic operations.
• LLMs+Graphs: Toward Graph-Native, Synergistic AI Systems — Integrating LLMs with graph computation creates synergistic AI systems that overcome LLM reasoning limitations.
• Neural-Symbolic Logic Query Answering in Non-Euclidean Space — Hyperbolic GNNs with learned curvature effectively capture hierarchical logic query structures for knowledge graph reasoning.
• GS-Quant: Granular Semantic and Generative Structural Quantization for Knowledge Graph Completion — GS-Quant aligns graph embeddings with LLM tokens using semantically coherent, hierarchically structured discrete codes.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.