Might be a dumb question, but what's the advantage of training agents on this data compared to just training a smaller chem model on the chemical screens and letting the agent use that model?
Replying to @allmeasures
We took huge chemical screens (millions of molecules) and told agents to find the best binder on a limited budget (they could only get results for 100 choices).
They buy a few at a time, get physical feedback, and update their hypotheses on binders.
5/10
