An LLM just got a side job as a lab assistant that keeps chemists honest.
Molecular optimization tools try to tweak a molecule's properties - make it bind better, dissolve better, whatever the target is - while making sure each change can actually be synthesized in a lab. The problem is combinatorial: every possible reaction and building block multiplies the search space, and it gets worse the more steps an optimization needs. A new system called MolReAct tackles this by using a tool-augmented large language model to propose a small, molecule-specific set of compatible reactions at each step, rather than searching the full universe of possibilities. It then applies a reinforcement learning method called Group Relative Policy Optimization, rewarded only at the end of the pathway, to make smarter multi-step decisions. Across 14 molecular optimization tasks, MolReAct posted the best Top-10 scores on 11 and the best sample efficiency on 12, beating existing baselines when the number of allowed test evaluations was limited.
This matters because "can we make this molecule better" and "can we actually synthesize it" have historically been separate problems, with a lot of computational drug and materials discovery producing molecules that look great on paper and are a nightmare to build. By having the LLM prune the reaction space at each step instead of brute-forcing it, MolReAct cuts the number of expensive oracle evaluations needed - the paper's real selling point for anyone paying for compute or lab time. Every output also comes with a template-grounded synthetic pathway, which is the part that turns a hypothetical molecule into a to-do list.
The bigger trend here is LLMs as narrowers, not deciders - the model isn't picking the final molecule, it's culling a huge search space down to something a classical optimizer can handle efficiently. That's a more modest and more defensible use of chemical "reasoning" than asking a model to just generate a drug candidate outright, and it's a pattern worth watching spread to other combinatorial design problems.