Reproduce “SafeSieve: From Heuristics to Experience in Progressive Pruning for LLM-based Multi-Agent Communication” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ruijia Zhang, Xinyan Zhao, Ruixiang Wang, Sigen Chen, Guibin Zhang, An Zhang, Kun Wang, Qingsong Wen
The SafeSieve framework introduces a progressive pruning method for LLM-based multi-agent systems, dynamically refining communication by integrating semantic heuristic initialization with accumulated historical performance feedback and 0-extension clustering. This approach improved average accuracy by up to 7.31 percentage points on complex tasks, reduced token consumption by 12.4%–27.8%, and enhanced robustness against prompt injection attacks.
Key insights
- A progressive strategy that integrates initial semantic heuristics with continuous, experience-driven feedback effectively optimizes multi-agent communication, mimicking a 'plan-then-adjust' paradigm.
- Structure-aware pruning via 0-extension clustering is crucial for maintaining network robustness, outperforming greedy pruning methods by preserving critical inter-agent complementarity.
- The framework provides a triple defense mechanism (preventive, responsive, structural) against prompt injection attacks, significantly enhancing system resilience.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.