Reproduce “A-SR: Self-Evolving Agentic LLMs for Symbolic Regression via Hierarchical Coordination” 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
Wenxiao Zhao, Dong Liu, Kaiyi Xu, Feng Liu, Zhen Zhao, Fei Ben, Shu Wang, Wenhao Li and 4 more
The A-SR framework introduces a self-evolving agentic LLM system for symbolic regression, which uses hierarchical coordination to route context-specific evidence to specialized LLM agents. This approach effectively addresses diverse failure modes and improves discovery reliability, achieving an average [email protected] of 48.30% on the LLM-SRBench scientific domains, an improvement over Deliberate Evolution's 25.79% with a Llama3.1-8B-Instruct backbone.
Key insights
- Recasting LLM-guided symbolic regression as a control problem centered on dynamic 'role–memory-view pairs' allows for more precise and context-specific guidance, effectively addressing heterogeneous search failures.
- A structured, hierarchical coordination system, integrating specialized LLM agents, adaptive protocol selection, and state-routed memory, is critical for robust and efficient navigation of the complex symbolic regression search space.
- Enabling self-evolution both within a single search run (via online adaptation) and across multiple runs (via trajectory distillation) empowers LLM agents to continuously adapt and improve their symbolic regression capabilities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.