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

2025-08-15

AI Summary

To address high communication redundancy, excessive token consumption, and fragmented optimization strategies (pre- vs. post-task) in LLM-based multi-agent systems, this paper proposes SafeSieve, a progressive adaptive pruning algorithm. Methodologically, SafeSieve integrates semantic-heuristic initialization with experience-driven refinement: it constructs an initial communication graph via LLM-based semantic evaluation, dynamically adjusts topology using performance feedback, and employs 0-extension clustering to preserve structural coherence—enabling smooth transition from prior rules to posterior optimization. Unlike static Top-k greedy pruning, SafeSieve supports dynamic, incremental topology reconstruction. Experiments across multiple benchmarks show SafeSieve achieves an average accuracy of 94.01%, reduces token consumption by 12.4%–27.8%, maintains robustness against prompt-injection attacks (accuracy drop only 1.23%), and lowers heterogeneous deployment costs by 13.3%.

Technology Category

• Multiagent Systems: Agent Communication
• Natural Language Processing: Safety and Robustness
• Search and Optimization: Learning to Search

Application Category

• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Search and Retrieval-Augmented AI: Agentic search
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

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 a robust, efficient, and scalable framework for practical multi-agent systems. Our code can be found in https://anonymous.4open.science/r/SafeSieve-D8F2FFUN.

Problem

• Reduces redundant communication in multi-agent systems
• Dynamically optimizes inter-agent token usage efficiently
• Enhances robustness against prompt injection attacks

Innovation

• Progressive adaptive pruning algorithm for multi-agent communication
• Dual-mechanism integrating semantic evaluation with performance feedback
• 0-extension clustering preserving coherent agent groups