SafeSieve: From Heuristics to Experience in Progressive Pruning for LLM-based Multi-Agent Communication

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Zhang, Ruijia, Zhao, Xinyan, Wang, Ruixiang, Chen, Sigen, Zhang, Guibin, Zhang, An, Wang, Kun, Wen, Qingsong
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866912981020835840
author Zhang, Ruijia
Zhao, Xinyan
Wang, Ruixiang
Chen, Sigen
Zhang, Guibin
Zhang, An
Wang, Kun
Wen, Qingsong
author_facet Zhang, Ruijia
Zhao, Xinyan
Wang, Ruixiang
Chen, Sigen
Zhang, Guibin
Zhang, An
Wang, Kun
Wen, Qingsong
contents LLM-based multi-agent systems exhibit strong collaborative capabilities but often suffer from redundant communication and excessive token overhead. Existing methods typically enhance efficiency through pretrained GNNs or greedy algorithms, but often isolate pre- and post-task optimization, lacking a unified strategy. To this end, we present SafeSieve, a progressive and adaptive multi-agent pruning algorithm that dynamically refines the inter-agent communication through a novel dual-mechanism. SafeSieve integrates initial LLM-based semantic evaluation with accumulated performance feedback, enabling a smooth transition from heuristic initialization to experience-driven refinement. Unlike existing greedy Top-k pruning methods, SafeSieve employs 0-extension clustering to preserve structurally coherent agent groups while eliminating ineffective links. Experiments across benchmarks (SVAMP, HumanEval, etc.) showcase that SafeSieve achieves 94.01% average accuracy while reducing token usage by 12.4%-27.8%. Results further demonstrate robustness under prompt injection attacks (1.23% average accuracy drop). In heterogeneous settings, SafeSieve reduces deployment costs by 13.3% while maintaining performance. These results establish SafeSieve as an efficient, GPU-free, and scalable framework for practical multi-agent systems. Our code can be found here: https://github.com/csgen/SafeSieve
format Preprint
id arxiv_https___arxiv_org_abs_2508_11733
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle SafeSieve: From Heuristics to Experience in Progressive Pruning for LLM-based Multi-Agent Communication
Zhang, Ruijia
Zhao, Xinyan
Wang, Ruixiang
Chen, Sigen
Zhang, Guibin
Zhang, An
Wang, Kun
Wen, Qingsong
Multiagent Systems
Artificial Intelligence
LLM-based multi-agent systems exhibit strong collaborative capabilities but often suffer from redundant communication and excessive token overhead. Existing methods typically enhance efficiency through pretrained GNNs or greedy algorithms, but often isolate pre- and post-task optimization, lacking a unified strategy. To this end, we present SafeSieve, a progressive and adaptive multi-agent pruning algorithm that dynamically refines the inter-agent communication through a novel dual-mechanism. SafeSieve integrates initial LLM-based semantic evaluation with accumulated performance feedback, enabling a smooth transition from heuristic initialization to experience-driven refinement. Unlike existing greedy Top-k pruning methods, SafeSieve employs 0-extension clustering to preserve structurally coherent agent groups while eliminating ineffective links. Experiments across benchmarks (SVAMP, HumanEval, etc.) showcase that SafeSieve achieves 94.01% average accuracy while reducing token usage by 12.4%-27.8%. Results further demonstrate robustness under prompt injection attacks (1.23% average accuracy drop). In heterogeneous settings, SafeSieve reduces deployment costs by 13.3% while maintaining performance. These results establish SafeSieve as an efficient, GPU-free, and scalable framework for practical multi-agent systems. Our code can be found here: https://github.com/csgen/SafeSieve
title SafeSieve: From Heuristics to Experience in Progressive Pruning for LLM-based Multi-Agent Communication
topic Multiagent Systems
Artificial Intelligence
url https://arxiv.org/abs/2508.11733