SafeSieve: From Heuristics to Experience in Progressive Pruning for LLM-based Multi-Agent Communication
Summary: arXiv:2508.11733v3 Announce Type: replace-cross
Recent advancements in large language model (LLM)-based multi-agent systems have demonstrated their potential for collaborative tasks. However, these systems frequently encounter challenges related to redundant communication and excessive token usage, which can hinder their overall efficiency. Traditional methods have sought to address these challenges using pretrained graph neural networks (GNNs) or greedy algorithms. Unfortunately, these approaches often treat pre- and post-task optimization as separate entities, leading to a lack of a cohesive optimization strategy.
Introducing SafeSieve
To tackle these limitations, researchers have introduced SafeSieve, a progressive and adaptive multi-agent pruning algorithm. SafeSieve is designed to dynamically refine inter-agent communication through an innovative dual-mechanism approach. This algorithm begins with an initial LLM-based semantic evaluation, which is subsequently enhanced by accumulated performance feedback. This combination allows for a seamless transition from heuristic initialization to experience-driven refinement.
Key Features of SafeSieve
- Progressive Pruning: SafeSieve utilizes a progressive pruning technique, allowing for continuous improvement in communication efficiency.
- 0-Extension Clustering: Unlike traditional greedy Top-k pruning methods, SafeSieve implements 0-extension clustering to maintain structurally coherent agent groups while effectively eliminating ineffective communication links.
- Dynamic Adaptation: The algorithm adapts based on performance feedback, ensuring that the pruning process evolves with the agents’ needs.
Performance Evaluation
Extensive experiments conducted across multiple benchmarks, including SVAMP and HumanEval, reveal that SafeSieve achieves an impressive average accuracy of 94.01%. Moreover, the algorithm significantly reduces token usage by 12.4% to 27.8%. These results highlight SafeSieve’s efficiency and effectiveness in optimizing multi-agent communication.
Robustness and Cost Efficiency
In addition to its performance metrics, SafeSieve has demonstrated robust resilience against prompt injection attacks, with only a 1.23% average accuracy drop. Furthermore, in heterogeneous environments, the algorithm successfully reduces deployment costs by 13.3% while maintaining comparable performance levels.
Conclusion
SafeSieve stands out as an efficient, GPU-free, and scalable framework tailored for practical multi-agent systems. By integrating experience-driven refinement with initial heuristic evaluations, SafeSieve establishes itself as a pivotal advancement in the field of LLM-based multi-agent communication. For those interested in exploring the implementation of this innovative algorithm, the code is available on GitHub: https://github.com/csgen/SafeSieve.