Reproduce “Resilient Consensus in Agentic AI” 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
Sribalaji C. Anand, George J. Pappas
Research from KTH and University of Pennsylvania applies classical resilient consensus theory to multi-agent large language model (LLM) systems, demonstrating that while LLM agents often fail to reach agreement under adversarial conditions, wrapping them with established resilient consensus filters can restore robust agreement, achieving approximately 100% consensus probability in complete graphs.
Key insights
- Prompted LLM agents, without external interventions, frequently fail to achieve consensus even in configurations where classical theory guarantees the existence of a robust algorithm.
- Integrating classical resilient consensus filters significantly improves the ability of LLM agents to reach agreement, recovering performance to levels expected by theoretical guarantees and sometimes exceeding them depending on the adversarial strategy.
- The efficacy of these classical filters is robust across different LLM temperatures and maximum interaction rounds, and they remain effective even when the precise number of Byzantine agents is misspecified.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.