HyAKE: Hybrid Adaptive Knowledge Expert for Knowledge Graph Reasoning

ACL ARR 2026 January Submission2569 Authors

Keywords: Knowledge Graph Reasoning, Hybrid Expert System, Adaptive Routing, Cost-Aware Inference, Large Language Models

Abstract: Knowledge-intensive NLP systems increasingly combine large language models (LLMs), knowledge graphs (KGs), and document retrieval. These sources have highly heterogeneous inference costs: LLM calls are accurate but expensive (120ms, cost 1.0), KG lookups are fast but limited (8ms, cost 0.02), and retrieval lies in between. Existing KG reasoning methods either rely on a single source or fuse multiple sources in a cost-blind manner, which can lead to suboptimal accuracy–cost trade-offs. We propose **HyAKE**, a cost-aware hybrid expert framework for KG reasoning. HyAKE integrates three specialists—a parametric LLM expert, a structural GNN expert over cached subgraphs, and a retrieval expert—with two key components: (i) a **Knowledge Graph Reasoning Planner (KGRP)** that decomposes complex queries into a DAG of sub-questions for dependency-aware, partially parallel execution; and (ii) an **Adaptive Knowledge Fusion Module (AKFM)** that performs query-specific, cost-aware expert routing with learned temperature networks and a CLUB-based decoupling loss to encourage complementary behaviors. Experiments on four benchmarks show that HyAKE improves MRR by 37–59% over strong baselines and by 10–19% over direct Qwen-2.5-7B prompting, while reducing normalized inference cost by 45% and achieving 2.2× lower latency. On unseen entities, HyAKE retains 79% of its transductive performance versus 45–51% for baselines, suggesting that gains are not solely due to memorization.

Paper Type: Long

Research Area: Mathematical, Symbolic, Neurosymbolic, and Logical Reasoning

Research Area Keywords: Information Retrieval and Text Mining, Machine Learning for NLP, Efficient/Low-Resource Methods for NLP, Language Modeling, Question Answering, NLP Applications

Contribution Types: NLP engineering experiment, Approaches to low-resource settings, Approaches low compute settings-efficiency

Languages Studied: English

Submission Number: 2569