Lower-performing language models repeatedly test minor chemical variations on low-scoring molecules instead of exiting unproductive neighborhoods