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

Jul 16, 2026

Upload by Asher

Multi-hop Question Answering over Knowledge Graphs suffers a key issue: retrieve-then-read pipelines lose differentiability, so the retriever cannot learn to cross the semantic gap when intermediate bridge entities have little or no lexical overlap with the query. RSF-GLLM decouples differentiable graph reasoning from answer generation using a Recurrent Soft-Flow module with a GRU-guided query updater and dynamic gating based on structural cues. Flow sparsity regularization supports convergence to discrete paths, which are textualized to fine-tune an LLM for topology-grounded generation. Experiments on WebQSP and CWQ show competitive results with improved inference efficiency.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages21

Reading Time53 min

Introduction

Semantic gap in KGQA and limitations of existing paradigms

Proposed RSF-GLLM framework and contributions

RSF-GLLM: Bridging the Semantic Gap in Multi-Hop Knowledge Graph QA via Recurrent Soft-Flow and Decoupled LLM Generation Sambaran Bandyopadhyay 1 Ananth Muppidi 2 arXiv :2607 .06527v 1 [ cs .CL] 7 Jul 2026 Abstract Multi-hop Question Answering over Knowledge Graphs faces a critical challenge: traditional retrieve-then-read pipelines break differentiability, preventing the retriever from learning to bridge the semantic gap where intermediate nodes lack lexical overlap with the query. To address this, we propose RSF-GLLM, a framework decoupling differentiable graph reasoning from answer generation. Our Recurrent Soft-Flow (RSF) module employs a GRU-guided query updater to propagate continuous relevance scores, utilizing a dynamic gating mechanism to traverse semantically dissimilar bridge nodes via structural cues. We introduce flow sparsity regularization to theoretically guarantee convergence from soft probabilities to discrete reasoning paths. These paths are extracted and textualized to fine-tune a Large Language Model (LLM), ensuring generation is grounded in factual topology. Experiments on WebQSPand CWQ demonstrate that RSF-GLLM achieves competitive performance with superior inference efficiency compared to LLM based computationally expensive approaches. 1. Introduction Reasoning over structured knowledge bases to answer complex, multi-hop queries is a fundamental challenge in artificial intelligence, essential for applications requiring high factual precision such as biomedical discovery and financial auditing (Wang et al., 2021) . While Large Language Models (LLMs) have demonstrated exceptional fluency in open-domain tasks (Brown et al., 2020), they frequently exhibit hallucinations and unfaithful reasoning when operating over long-tail knowledge or complex logical chains (Ji et al., 2023 ; Nandy & Bandyopadhyay, 2025) . Conse- 1Adobe Research 2Adobe Systems. Correspondence to: Sambaran Bandyopadhyay <[samb.bandyo@gmail.com](samb.bandyo@gmail.com) >. Proceedings of the 43 rd International Conference on Machine Learning, Seoul, South Korea. PMLR 306, 2026 . Copyright 2026 by the author(s) . quently, Knowledge Graphs (KGs) remain indispensable for grounding generation in verifiable facts (Hogan et al., 2021 ; Nandy & Bandyopadhyay, 2025) . However, effective retrieval over KGs is impeded by the semantic gap. Consider the query: “What awards did the director of Inception win?”. To answer this, a system must traverse from the topic entity “Inception” to the intermediate node “Christopher Nolan”(the director), and finally to “Academy Award”. Crucially, the intermediate bridge entity “Christopher Nolan”shares no lexical overlap with the query terms “awards” or“win”, rendering standard similarity-based traversal ineffective (Sun et al., 2018 ; Xiong et al., 2017) . Current approaches to Knowledge Graph Question Answering (KGQA) effectively bifurcate into two paradigms, each utilizing distinct inductive biases yet suffering from complementary limitations. The first category comprises iterative subgraph retrieval methods, such as GraftNet (Sun et al., 2018), PullNet (Sun et al., 2019), and NSM (He et al., 2021) . These architectures typically treat reasoning as a discrete Retrieve-then-Read process (Karpukhin et al., 2020) . However, the discrete selection of nodes at each hop breaksend-to-end differentiability, preventing the retriever from adapting to downstream errors (Kim et al., 2023) . Furthermore, rigid entity-linking requirements often cause these models to fail when the requisite bridge nodes are semantically dissimilar to the query text (Liang et al., 2024) . The second, more recent paradigm integrates the generative capabilities of LLMs directly. Frameworks such as Graph of Thoughts (Besta et al., 2024) and RoG (Luo et al., 2024) attempt to improve expressivity by structuring LLM reasoning as an iterative graph traversal. However, these agentic approaches incur prohibitive computational costs, of

FAQs

What problem does RSF-GLLM address in multi-hop knowledge graph question answering?

RSF-GLLM targets the semantic gap where differentiable retrieve-then-read pipelines cannot learn to bridge intermediate nodes that lack lexical overlap with the query.

How does RSF-GLLM decouple reasoning from answer generation?

It separates a differentiable Recurrent Soft-Flow reasoning module from downstream LLM answer generation, extracting and textualizing the reasoning paths to ground generation in graph topology.

What mechanisms help RSF-GLLM traverse semantically dissimilar bridge entities?

A GRU-guided query updater and a dynamic gating mechanism modulate the balance between node content and graph structure, allowing traversal via structural relations even when semantic similarity is weak.

Jul 16, 2026

Upload by Asher