In materials research, this may be especially well suited to problems with delayed reward signals, such as synthesis planning, reaction-pathway search, and analysis procedures. It may be tried by retrieving step-by-step trajectories that include past experimental notes, simulation logs, and even failed conditions, and using them to generate the next action candidates or as evidence for policy updates.
At first, it is safer not to fully automate the agent itself, but to make 50~200 trajectories already organized by a human searchable and apply them only to a small policy model. After that, one could evaluate whether it “repeats similar failures,” “converges more quickly to successful paths,” and “increases exploration of new combinations.”