Mitigating Semantic Drift in Multi-Agent Communication: A Dynamic Neuro-Symbolic Approach

doi:10.12142/ZTECOM.202602008

ZTE Communications ›› 2026, Vol. 24 ›› Issue (2): 64-70.DOI: 10.12142/ZTECOM.202602008

• Special Topic • Previous Articles Next Articles

Mitigating Semantic Drift in Multi-Agent Communication: A Dynamic Neuro-Symbolic Approach

Xie Linhao¹^,², Li Fan³, Wu Mingxuan⁴, Song Yong¹(), Ouyang Ye¹

^1.AsiaInfo Technologies, Beijing 100193, China
^2.Communication University of China, Beijing 100024, China
^3.China Unicom Beijing Branch Network Optimization Center, Beijing 100032, China
^4.CRSC Information Industry Co. , Ltd. , Beijing 100070, China

Received:2026-02-24 Online:2026-06-16 Published:2026-06-16
About author:Xie Linhao is currently an undergraduate student at Communication University of China. He is also an intern at the Product R&D Center of AsiaInfo Technologies. During his university years, he received several awards, including the First Prize in the 2025 AI.Talk National College Student Artificial Intelligence Knowledge Competition, the Second Prize in the North China region of the 15th MathorCup Mathematical Application Challenge in 2025, and the National Third Prize in the 2024 National Mathematics Competition. His research interests include multi-agent systems, computer vision, and intelligent robotics.
Li Fan is a Senior Engineer at China Unicom Beijing Branch. She holds a Master of Engineering from Beijing University of Posts and Telecommunications, China. Her research interests include 4G/5G network optimization, mobile network digital operation, intelligentization of mobile communication networks, intelligent optimization and operation and maintenance, and emerging mobile communication technologies.
Wu Mingxuan received his MS degree from Peking University, China. He is currently the Deputy Manager of the Information Technology Department at CRSC Information Industry Co., Ltd., and holds the professional title of Engineer. His research interests include application of AI agents in vertical fields, digital transformation for large enterprises, and smart city solutions.
Song Yong (songyong@asiainfo.com) received his MS degree from Peking University, China. He is currently an algorithm expert at the Product R&D Center of AsiaInfo Technologies. He received the Second Prize of the Beijing Science and Technology Progress Award. His research interests include automatic ontology generation, multi-agent systems, and the Internet of Agents.
Ouyang Ye is a professor and an IEEE Fellow. He serves as the Chief Executive Officer and Chief Technology Officer at AsiaInfo Technologies. He holds a Bachelor of Engineering from Southeast University in China, a Master of Science from Tufts University in the United States, a second Master of Science from Columbia University in the United States, and a PhD from Stevens Institute of Technology in the United States. He has extensive experience in large-scale team management and R&D innovation in the ICT field. He focuses on cross-domain innovation and the commercialization of technologies in cellular networks, AI, and data science.
Supported by:
Mobile Information Networks?National Science and Technology Major Project of China(2025ZD1304800)

Abstract

Abstract:

The emergence of multi-agent systems (MAS) based on large language models (LLMs) has enabled autonomous collaboration on complex, goal-oriented tasks. However, effective interaction is frequently hindered by semantic drift, a phenomenon where heterogeneous agents assign conflicting meanings to shared terminology due to differing internal prompts or domain knowledge. Existing communication paradigms either rely on unconstrained natural language, which suffers from structural vagueness, or rigid symbolic schemas that fail to adapt to emergent concepts. To address this gap, we propose DOA, a novel dynamic ontology alignment framework that serves as a semantic mediation layer for MAS. DOA integrates a proactive semantic prober to detect conceptual mismatches and a neuro-symbolic aligner that reconciles local semantic structures in real time. By grounding fluid natural language dialogues in an evolving shared ontology, our framework ensures deterministic mutual understanding over long-horizon tasks. Empirical evaluations in cross-domain supply chain and healthcare coordination scenarios demonstrate that DOA improves task success rates by an average of 31.5% and reduces communication overhead (token consumption) by 50% compared to state-of-the-art baselines. Our results provide a robust and scalable foundation for semantic consistency in next-generation industrial-grade AI systems.

Key words: DOA, multi-agent systems, neuro-symbolic AI, semantic drift, large language models

Xie Linhao, Li Fan, Wu Mingxuan, Song Yong, Ouyang Ye. Mitigating Semantic Drift in Multi-Agent Communication: A Dynamic Neuro-Symbolic Approach[J]. ZTE Communications, 2026, 24(2): 64-70.

Figures/Tables 5

Figure 1 Semantic drift in multi-agent communication

Table 1 Task success rate across frameworks in CDSC and AHC

Framework	CDSC/%	AHC/%	Average/%	Improvement over Baseline/%
Raw-NL	62.5	54.0	58.3	31.5
Fixed-Schema	45.0	38.5	41.8	48.0
LLM-Prompt-Only	67.5	61.5	64.5	25.3
DOA (Ours)	91.5	88.0	89.8	-

Table 2 System success rate vs. semantic drift (δ)

Drift ( $δ$ )	Raw-NL/%	Fixed-Schema/%	LLM-Prompt-Only/%	DOA (Ours)/%
0.1 (Low)	90.3	88.0	92.5	96.0
0.2	75.5	72.8	83.3	94.0
0.3 (Threshold)	58.3	41.8	64.5	89.8
0.4	32.3	22.0	43.5	88.3
0.5 (High)	16.5	12.3	24.3	82.5
0.6 (Critical)	6.5	5.0	11.0	65.8

Table 2 System success rate vs. semantic drift (δ)

Drift ( $δ$ )	Raw-NL/%	Fixed-Schema/%	LLM-Prompt-Only/%	DOA (Ours)/%
0.1 (Low)	90.3	88.0	92.5	96.0
0.2	75.5	72.8	83.3	94.0
0.3 (Threshold)	58.3	41.8	64.5	89.8
0.4	32.3	22.0	43.5	88.3
0.5 (High)	16.5	12.3	24.3	82.5
0.6 (Critical)	6.5	5.0	11.0	65.8

Figure 2 Robustness analysis of the impact of semantic drift δ on task success rate, where the DOA framework maintains stability compared to the sharp decline of baseline methods

Table 3 Communication overhead analysis: comparison of efficiency metrics per successful task completion

Metric	Raw-NL	Fixed-Schema	LLM-Prompt-Only	DOA (Ours)
Avg. turns to consensus	8.4	2.0^*	6.2	3.1
Avg. tokens per task	4 200	1 200	3 850	1 950
Alignment latency / s	N/A	0.3	N/A	1.4

References 23

[1]	Qian C, Liu W, Liu H Z, et al. ChatDev: communicative agents for software development [C]//Proc. 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). ACL, 2024: 15174–15186. DOI: 10.18653/v1/2024.acl-long.810
[2]	Xi Z H, Chen W X, Guo X, et al. The rise and potential of large language model based agents: a survey [J]. Science China information sciences, 2025, 68(2): 121101. DOI: 10.1007/s11432-024-4222-0
[3]	Yang Y X, Chai H C, Shao S, et al. AgentNet: decentralized evolutionary coordination for LLM-based multi-agent systems [PP/OL]. V1. arXiv (2025-04-01)[2025-12-12].
[4]	Park J S, O'Brien J, Cai C J, et al. Generative agents: interactive simulacra of human behavior [C]//36th Annual ACM Symposium on User Interface Software and Technology. ACM, 2023: 1–22. DOI: 10.1145/3586183.3606763
[5]	Becker J, Kaesberg L B, Stephan A, et al. Stay focused: problem drift in multi-agent debate [PP/OL]. V1. arXiv (2025-02-26)[2025-12-12].
[6]	Chen W Z, Su Y S, Zuo J W, et al. AgentVerse: facilitating multi-agent collaboration and exploring emergent behaviors [C]//International Conference on Learning Representations. ICLR, 2024: 1–43. DOI: 10.48550/arXiv.2308.10848
[7]	Noy N F, McGuinness D L. Ontology development 101: a guide to creating your first ontology [EB/OL]. [2026-01-13].
[8]	Giglou H B, D'Souza J, Engel F, et al. LLMs4OM: matching ontologies with large language models [PP/OL]. V2. arXiv (2024-04-23)[2026-01-14].
[9]	Qiang Z C, Wang W Q, Taylor K. Agent-OM: leveraging LLM agents for ontology matching [J]. Proceedings of the VLDB endowment, 2024, 18(3): 516–529. DOI: 10.14778/3712221.3712222
[10]	Hitzler P, Sarker M K. Neuro-symbolic artificial intelligence: the state of the art [M]. Amsterdam: IOS Press, 2021. DOI: 10.3233/faia342
[11]	Chang E Y. Multi-LLM agent collaborative intelligence: the path to artificial general intelligence [M]. New York: ACM, 2025. DOI: 10.1145/3749421
[12]	Hong S R, Zhuge M C, Chen J, et al. MetaGPT: meta programming for a multi-agent collaborative framework [C]//International Conference on Learning Representations. ICLR, 2024: 1–29. DOI: 10.48550/arXiv.2308.00352
[13]	Cancedda N, Dessi R, Dwivedi-Yu J, et al. Toolformer: language models can teach themselves to use tools [C]//Proceedings of Advances in Neural Information Processing Systems 36. Neural Information Processing Systems Foundation, Inc. (NeurIPS), 2023: 68539–68551. DOI: 10.52202/075280-2997
[14]	Renney H, Nethercott M N, Renney N, et al. LLM-enabled multi-agent systems: empirical evaluation and insights into emerging design patterns & paradigms [PP/OL]. V1. arXiv (2026-01-06)[2026-02-01].
[15]	Wang Z, Moriyama S, Wang W Y, et al. Talk structurally, act hierarchically: a collaborative refinement framework for LLM multi-agent systems [J]. IEEE transactions on artificial intelligence, 2026: 1–11. DOI: 10.1109/tai.2026.3676025
[16]	Jiménez-Ruiz E, Cuenca Grau B. LogMap: logic-based and scalable ontology matching [C]//The Semantic Web―ISWC 2011. Springer, 2011: 273–288. DOI: 10.1007/978-3-642-25073-6_18
[17]	He Y, Chen J Y, Dong H, et al. Exploring large language models for ontology alignment [PP/OL]. V1. arXiv (2023-09-12)[2026-01-18].
[18]	Kamali D, Barezi E J, Kordjamshidi P. NeSyCoCo: a neuro-symbolic concept composer for compositional generalization [J]. Proceedings of the AAAI conference on artificial intelligence, 2025, 39(4): 4184–4193. DOI: 10.1609/aaai.v39i4.32439
[19]	Schelble B G, Flathmann C, McNeese N J, et al. Let's think together! assessing shared mental models, performance, and trust in human-agent teams [J]. Proceedings of the ACM on human-computer interaction, 2022, 6: 1–29. DOI: 10.1145/3492832
[20]	He J D, Treude C, Lo D. LLM-based multi-agent systems for software engineering: literature review, vision, and the road ahead [J]. ACM transactions on software engineering and methodology, 2025, 34(5): 1–30. DOI: 10.1145/3712003
[21]	Bougzime O, Jabbar S, Cruz C, et al. Unlocking the potential of generative AI through neuro-symbolic architectures: benefits and limitations [PP/OL]. V1. arXiv (2025-02-16)[2026-02-23].
[22]	El-Sappagh S, Franda F, Ali F, et al. SNOMED CT standard ontology based on the ontology for general medical science [J]. BMC medical informatics and decision making, 2018, 18: 76. DOI: 10.1186/s12911-018-0651-5
[23]	Möller M, Sintek M, Biedert R, et al. Representing the international classification of diseases version 10 in OWL [C]//International Conference on Knowledge Engineering and Ontology Development. KEOD, 2010: 50–59

Mitigating Semantic Drift in Multi-Agent Communication: A Dynamic Neuro-Symbolic Approach

RichHTML

PDF

Knowledge

Abstract

Cite this article

share this article

Figures/Tables 5

References 23

Related Articles 1

Recommended Articles

Metrics