ReMedI: Inference-Time Refinement and Self-Distillation for Controllable Molecular Optimization

ReMedI builds on a molecular language model to propose molecular edits and uses Metropolis-Hastings search for multi-objective optimization over property improvement, property preservation, and lead similarity. On the C-MuMOInstruct and SMDD benchmarks, refinement lets frozen base models across three backbone architectures achieve significantly higher multi-objective optimization success rates.