MobileKGQA: On-Device KGQA System on Dynamic Mobile Environments

ICLR 2026 Conference SubmissionAnonymous Authors
Knowledge Graph Question AnsweringLarge Language Model
Abstract:

Developing a mobile system capable of generating responses based on stored user data is a crucial challenge. Since user data is stored in the form of Knowledge Graphs, the field of knowledge graph question answering (KGQA) presents a promising avenue towards addressing this problem. However, existing KGQA systems face two critical limitations that preclude their on-device deployment: resource constraints and the inability to handle data accumulation. Therefore, we propose MobileKGQA, the first on-device KGQA system capable of adapting to evolving databases with minimal resource demands. MobileKGQA significantly reduces computational overhead through embedding hashing. Moreover, it successfully adapts to evolving databases under resource constraints through a novel annotation generation method. Its mobile applicability is validated on the NVIDIA Jetson Orin Nano edge-device platform, achieving 20.3% higher performance while using only 30.4% of the energy consumed by the SOTA (state of the art). On standard KGQA benchmarks, using just 7.2% of the computation and 9% of the parameters, MobileKGQA demonstrates performance that is empirically indistinguishable from the SOTA and outperforms baselines under distribution shift scenarios.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Overall Novelty Assessment

The paper proposes MobileKGQA, an on-device KGQA system designed to handle evolving knowledge graphs with minimal computational overhead. Within the taxonomy, it occupies the 'Dynamic Knowledge Graph Adaptation' leaf under 'On-Device KGQA System Optimization'. Notably, this leaf contains only the original paper itself—no sibling papers were identified in the taxonomy. This suggests the specific combination of on-device execution, dynamic database adaptation, and resource-constrained KGQA represents a relatively sparse research direction within the broader field of mobile knowledge graph systems.

The taxonomy reveals neighboring research directions that address mobile KGQA through alternative strategies. The sibling leaf 'Latency and Memory Optimization for Mobile QA' focuses on deep learning model optimizations without knowledge graph structures, while the 'Cloud-Edge Collaborative Intelligence' branch explores hybrid architectures that partition computation between devices and servers. Domain-specific approaches in the third branch tailor KGQA to particular applications like geoscience surveys or personal knowledge management. MobileKGQA diverges from these by prioritizing fully autonomous on-device operation with dynamic graph handling, rather than cloud offloading or domain-specific constraints.

Among the three contributions analyzed, the embedding hashing module shows the most substantial prior work overlap. The literature search examined ten candidates for this contribution, identifying one that appears to provide refutable prior work. In contrast, the system-level contribution (first on-device KGQA with hashing-based retrieval) examined two candidates with no clear refutations, and the annotation generation method examined three candidates, also without refutations. These statistics reflect a limited search scope of fifteen total candidates, suggesting the analysis captures high-relevance matches but may not represent exhaustive coverage of the broader KGQA and mobile AI literature.

Based on the limited search scope, the work appears to occupy a relatively unexplored intersection of on-device execution, dynamic knowledge graph adaptation, and resource efficiency. The embedding hashing component shows connections to existing techniques, while the system-level integration and annotation generation method appear less directly addressed in the examined candidates. The sparse taxonomy leaf and modest candidate pool suggest caution in drawing definitive conclusions about novelty without broader literature coverage.

Taxonomy

Core-task Taxonomy Papers
6
3
Claimed Contributions
15
Contribution Candidate Papers Compared
1
Refutable Paper

Research Landscape Overview

Core task: knowledge graph question answering on resource-constrained mobile devices. The field addresses the challenge of deploying sophisticated KGQA systems in environments with limited computational power, memory, and energy. The taxonomy reveals four main branches that capture distinct strategies for overcoming these constraints. On-Device KGQA System Optimization focuses on techniques that enable efficient local execution through model compression, dynamic knowledge graph adaptation, and runtime optimization. Cloud-Edge Collaborative Intelligence explores hybrid architectures that distribute computation between mobile devices and remote servers, balancing latency and resource usage as seen in works like Mobile Edge QA[4] and Cloud Edge Entity Linking[5]. Domain-Specific Lightweight KGQA Applications tailor solutions to particular use cases such as personal knowledge management (Personal Knowledge QA[2]) or specialized domains (Mineral Survey QA[1]), leveraging domain constraints to reduce complexity. Finally, IoT Security and Forensics Infrastructure addresses the unique requirements of secure, auditable KGQA in IoT ecosystems, exemplified by Blockchain IoT Forensics[6]. A central tension across these branches involves the trade-off between autonomy and performance: fully on-device solutions offer privacy and offline capability but face severe resource limits, while cloud-edge approaches can leverage powerful remote models at the cost of latency and connectivity dependence. Within the On-Device KGQA System Optimization branch, MobileKGQA[0] emphasizes dynamic knowledge graph adaptation, allowing the system to selectively load and update graph segments based on query patterns and device state. This approach contrasts with earlier efforts like Deqa[3], which focused more on static model compression, and differs from collaborative strategies such as Mobile Edge QA[4] that offload heavy reasoning to edge servers. MobileKGQA[0] thus represents a middle path, seeking to maximize on-device intelligence through adaptive resource management rather than relying primarily on external computation or fixed lightweight models.

Claimed Contributions

MobileKGQA: First on-device KGQA system with hashing-based retrieval

The authors introduce MobileKGQA, the first knowledge graph question answering system designed for on-device deployment and training. It uses a hashing module to compress embeddings into binary codes and a reasoning module for efficient retrieval, enabling adaptation to evolving databases under resource constraints.

2 retrieved papers
Embedding hashing module with mutual information maximization

The system employs a hashing module that transforms high-dimensional floating-point embeddings into low-dimensional binary hash codes while preserving semantic information by maximizing mutual information. This approach eliminates the need to store gigabyte-scale embeddings or regenerate them during training.

10 retrieved papers
Can Refute
Sequential reasoning-based annotation generation method

The authors propose a stepwise annotation generation process that incrementally integrates structured knowledge to construct logically coherent questions. This method decomposes complex reasoning into simpler steps, reducing token generation requirements and enabling supervised training without data leakage in resource-constrained mobile settings.

3 retrieved papers

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Within the taxonomy built over the current TopK core-task papers, the original paper is assigned to a leaf with no direct siblings and no cousin branches under the same grandparent topic. In this retrieved landscape, it appears structurally isolated, which is one partial signal of novelty, but still constrained by search coverage and taxonomy granularity.

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

MobileKGQA: First on-device KGQA system with hashing-based retrieval

The authors introduce MobileKGQA, the first knowledge graph question answering system designed for on-device deployment and training. It uses a hashing module to compress embeddings into binary codes and a reasoning module for efficient retrieval, enabling adaptation to evolving databases under resource constraints.

Contribution

Embedding hashing module with mutual information maximization

The system employs a hashing module that transforms high-dimensional floating-point embeddings into low-dimensional binary hash codes while preserving semantic information by maximizing mutual information. This approach eliminates the need to store gigabyte-scale embeddings or regenerate them during training.

Contribution

Sequential reasoning-based annotation generation method

The authors propose a stepwise annotation generation process that incrementally integrates structured knowledge to construct logically coherent questions. This method decomposes complex reasoning into simpler steps, reducing token generation requirements and enabling supervised training without data leakage in resource-constrained mobile settings.