Reproduce “AgentConductor: Topology Evolution for Multi-Agent Competition-Level Code Generation” 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
Siyu Wang, Ruotian Lu, Zhihao Yang, Yuchao Wang, Yanzhou Zhang, Lei Xu, Qimin Xu, Guojun Yin and 2 more
AgentConductor introduces a multi-agent system where an LLM orchestrator dynamically generates and refines interaction topologies for competition-level code generation. This system achieves top accuracy on benchmarks while reducing token consumption through adaptable, difficulty-aware agent collaborations.
Key insights
- Supervised Fine-Tuning (SFT) is crucial for an LLM orchestrator to acquire foundational knowledge for generating syntactically and semantically correct multi-agent interaction graphs.
- Optimizing dynamic topology generation with Reinforcement Learning, guided by a reward function balancing code performance and interaction density, yields superior efficiency and adaptability.
- Layered Directed Acyclic Graphs offer an expressive yet efficient framework for agent collaboration, enabling both parallel processing and flexible cross-layer dependencies vital for complex problem-solving.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.