RetroAgent Uses LLMs for Advanced Retrosynthesis Planning

Multi-step retrosynthesis planning, which involves breaking down a target molecule into readily available building blocks through a series of chemical reactions, is a highly complex task even for expert chemists due to its vast combinatorial search space. Traditional computational methods often combine tree search with value networks that evaluate candidates in isolation, lacking a holistic view of complete multi-step routes. While large language models (LLMs) have recently been applied to this task, their interfaces have typically limited comprehensive exploration. To overcome these limitations, researchers introduce RetroAgent, an LLM-powered agent that bridges symbolic search and neural reasoning. It achieves this through a specialized harness equipped with structured memory. This memory allows RetroAgent to observe the entire search state, including previously explored routes, alternative pathways, and properties of intermediate compounds. By grounding its decisions in both global progress and specific domain knowledge, RetroAgent demonstrates strong performance and generalization capabilities on both in-distribution and out-of-distribution benchmarks. This approach significantly enhances the agent's ability to make informed decisions for complex retrosynthesis planning.