A new set of energy values to predict the secondary structures in RNA molecules has been derived through a multiple-step refinement procedure. It achieves more than 80% success in predicting the cloverleaf pattern in tRNA (200 sequences tested) and more than 60% success in predicting the consensus folding of 5S RNA (100 sequences). Improvements in our initial program for predicting secondary structures, based on the principle of the ‘incompatibility islets’ made possible the work on 5S RNA. The program was speeded up by introducing a dynamic grouping of the islets into three disjoint blocks. The novel features in the energy model include i) an evaluation of the contribution of odd pairs according to their position within a segment ii) a penalty for internal loops related to their dissymmetry iii) a bonus for bulge loops when the two terminal paired bases at the junction point are both pyrimidines.