Abstract
Rapid, sustainable redesign of large functional molecules demands efficient exploration of vast chemical spaces. Chemical language models (CLMs), especially transformers, can learn long-range structure-property relationships and enable fast candidate generation after training. However, inverse molecular design is ill-posed - many structures can meet the same target - and conditioned generation often decodes to invalid or off-spec molecules. To address this challenge, we propose a CLM-based inverse design framework that optimises latent representations toward target properties and explicitly evaluates round-trip fidelity, i.e., whether decoded candidates remain on-target after decoding and forward re-evaluation. To improve reliability, we introduce post-decoding beam re-ranking using round-trip consistency and a predictor-guided minimal-edit repair step that corrects invalid near-misses while preserving closeness to the target property. We demonstrate the approach on surfactant critical micelle concentration (CMC) design, benchmarking existing large pretrained CLMs against our lightweight domain-trained CLM. The framework produces a high proportion of valid and diverse molecules (~90%) while maintaining target property error near 1%. Moreover, atom-level saliency analysis confirms that the generated structures follow established surfactant design rules, supporting interpretable structure-property control. Overall, the framework provides an efficient and broadly applicable solution to reliable inverse design of novel functional molecules.