What kind of system you have? Finding a transition state might be difficult and it usually requires knowledge of the structure. Using the QST2 method, you can input the coordinates of initial and final states to try a sort of "blind" TS search. However, it is easier if you know what the structure should look like, as you can slightly perturb the initial state structure along some reaction coordinate to approximate the TS geometry. Additional information could help us in providing further recommendations, as each case is different...

Thanks for your answer. The system contains up to 180 atoms with significant π-π interactions, it is till of doubt wether the computer facility could afford such large system using QST2 to locate TS. I am sorry I could not provide the detailed molecular structure, since those are related to unpublished work.

The molecular model competes in triplet and quintet state, and fortunately we have got the TSs on the quintet PES, while failed on some cases of the triplet PES. For those faied cases, the first two convergency criteria can  be easily reached and the eigenvector is as expected. 

So I thought the problem maybe related to the opt algorithnm of Gaussian software, or something beyond my knowledge. 

try using a smaller step

e.g. MaxStep=10