Closing the Prior-Posterior Loop: Self-Reflective Molecular Design with Analysis-Driven LLM Iteration

The paper presents a novel approach to molecular design using a large language model (LLM) that integrates a self-reflection module with retrieval-augmented generation (RAG). By incorporating detailed physicochemical data such as orbital energies and electron densities instead of scalar feedback, the system achieves a deviation of 0.0014 eV in predicting HOMO-LUMO gaps, demonstrating a 100% success rate and outperforming traditional methods. This advancement enables more mechanistic and informed iterative design processes, significantly enhancing the potential for precise molecular engineering in various applications.