Reproduce “HiVA: Self-organized Hierarchical Variable Agent via Goal-driven Semantic-Topological Evolution” 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
Jinzhou Tang, Jusheng Zhang, Qinhan Lv, Sidi Liu, Jing Yang, Chengpei Tang, Keze Wang
HiVA introduces a framework for self-organized hierarchical variable agents that co-evolves both agent semantics and collaborative topology from a single initial agent, utilizing "Textual Gradients" for optimization in discrete spaces. The system achieves an average accuracy of 89.2% across diverse benchmarks and significantly improves cost-efficiency in complex agentic environments like GAIA.
Key insights
- Co-evolving agent semantics (what agents do) and collaborative topology (how they organize) from a singleton agent is crucial for scalable, adaptive, and self-organized intelligence.
- Textual Gradients can effectively serve as a functional substitute for backpropagation in discrete, non-differentiable problem spaces, enabling LLM-driven optimization of multi-agent systems.
- Leveraging the evolving multi-agent graph structure itself as a hierarchical memory allows the system to utilize historical experience at macro, meso, and micro levels for continuous self-improvement.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.