why LIBSVM runs so slowly for 0/1-based feature matrix

I am using Libsvm to perform text classification tasks. I normally uses binary occurrence, TF/IDF to build feature set for the input documents. It normally takes quite longer for Libsvm to finish training process when using binary occurrence. I doubt that the underlying solver of Libsvm has some difficulties to get converged for the 0/1 matrix resulted from the binary occurrence-based feature construction. Are there any reference or formal explanations on this kind of observation?