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

2026-06-08

AI Summary

This work addresses the limitations of conventional large language models (LLMs) in molecular design, which rely solely on scalar feedback and resort to trial-and-error generation without insight into failure mechanisms. For the first time, the authors integrate physicochemical mechanistic information—such as orbital energies, atomic charges, and electron densities derived from first-principles calculations—into the LLM design loop. They propose a causal reasoning framework that combines retrieval-augmented generation with a self-reflection mechanism, transforming the model from a stochastic sampler into an interpretable causal reasoner. Evaluated on HOMO–LUMO gap design tasks targeting 1.0–5.0 eV, the method achieves a remarkable bias as low as 0.0003 eV and 100% success on medium-difficulty tasks, substantially outperforming scalar-feedback baselines. Furthermore, it generalizes successfully to dipole moment design, demonstrating the effectiveness and broad applicability of mechanism-driven iterative optimization.

Abstract

Can a general-purpose large language model design molecules with the precision of a seasoned chemist? Current LLM-based frameworks answer this question with scalar feedback loops-generate, score, reject-that amount to informed trial-and-error. Here we show that replacing a single number with the full physicochemical rationale from first-principles calculations transforms the LLM from a stochastic sampler into a causal reasoner. Our system couples retrieval-augmented generation with a self-reflection module that feeds orbital energies, atomic charges, and electron densities-rather than compressed scores-back into the design loop. On HOMO-LUMO gap targets from 1.0 to 5.0 eV, this structure-property-relationship (SPR) reflection achieves a deviation as low as 0.0003 eV and a 100% success rate on moderate tasks, decisively outperforming scalar-feedback and non-reflective baselines. The framework generalizes seamlessly to dipole-moment design and proves robust across five distinct LLM backbones. These results establish a new paradigm: when the model understands not only that a molecule fails, but why, iterative molecular design becomes genuinely mechanistic.

Problem

• molecular design
• large language model
• structure-property relationship
• self-reflection
• iterative design

Innovation

• self-reflection
• structure-property relationship
• first-principles feedback
• retrieval-augmented generation
• causal reasoning