Molecular optimization stops at one turn: MFG trains an LLM on multi-turn feedback and reaches 81.50% average similarity-constrained success across ten tasks

A molecular optimization model is now trained to improve the same molecule across several turns instead of delivering one generation and stopping. Earlier approaches split in two: some wrote the previous turn's property feedback into the prompt at inference but never trained the model to use it, while others used supervised fine-tuning or reinforcement learning limited to non-iterative, single-turn generation. On ten multi-objective optimization tasks in MuMOInstruct, GeLLM3O-MFG, built on GeLLM3O, reaches an average similarity-constrained success rate of 81.50%, above the strongest baseline BioMedGPT-Mol and above the base model GeLLM3O.

At every turn the model receives a feedback vector recording whether each target property improved or degraded relative to the source molecule, together with the structural similarity between the current molecule and the source. The training reward is aligned with that vector: below the similarity threshold only a similarity-dependent penalty applies, so large property gains cannot offset structural drift; above the threshold the reward combines normalized property improvement, similarity, and progress over the best molecule from earlier turns. Training uses group-relative policy optimization (GRPO, a reinforcement-learning method that updates the policy from the relative scores of candidates in a group), sampling several candidate molecules per turn and carrying one intermediate molecule forward under a hierarchical selection criterion; evaluation covers five in-distribution and five out-of-distribution MuMOInstruct tasks.

Evaluation is confined to the MuMOInstruct dataset, covering combinations of six properties: blood-brain barrier permeability, dopamine D2 receptor binding affinity, human intestinal absorption, mutagenicity, penalized logP and drug-likeness, split into five in-distribution and five out-of-distribution tasks; property improvements come from property oracles and have not been experimentally validated.

Multi-turn feedback depends on property oracles, whose accuracy sets feedback reliability, and the property improvements seen in evaluation have not been experimentally validated; the training reward normalizes with reference optimized molecules supplied by MuMOInstruct, so datasets without such target molecules may be hard to use directly; the method is instantiated only on GeLLM3O, and whether it works as well for other molecular optimization LLMs still needs evaluation; objectives beyond the six properties remain unexplored.