Reproduce “LARC: Towards Human-level Constrained Retrosynthesis Planning through an Agentic Framework” 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

Frazier N. Baker, Daniel Adu-Ampratwum, Reza Averly, Botao Yu, Huan Sun, Xia Ning

LARC, an LLM-based agentic framework, integrates practical constraints directly into retrosynthesis planning to generate synthetic routes that avoid hazardous substances. It achieves a 72.9% success rate across various safety-constrained tasks and plans routes over four times faster than human experts in certain scenarios.

Key insights

  • Integrating LLM agents with specialized tools (Agent-as-a-Judge) directly into the planning loop is effective for solving complex, constrained scientific problems like retrosynthesis.
  • Modifying the value function of existing unconstrained retrosynthesis planners to include constraint satisfaction scores enables proactive, dynamic guidance in generating compliant synthetic routes.
  • Smaller, cost-effective open-source LLMs can achieve high performance in specialized scientific tasks when effectively grounded with domain-specific tools.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.