In Leash Bio's benchmark experiments, certain language model agents selected molecules that contradicted the agents' own stated structure-activity relationship hypotheses