GRMM is a toolkit for performing inference and learning in graphical models of arbitrary structure. Its main features are:
GRMM supports arbitrary factor graphs, which subsume both Markov random fields and Bayesian networks. It includes efficient implementations of several inference algorithms, including junction tree, belief propagation, and Gibbs sampling. All inference algorithms work for factors of any size (not just pairwise). [Developer's Guide]
Ability to train conditional random fields of arbitrary structure and parameter tying. [Quick Start]
GRMM can handle continuous features, but only discrete variables.
GRMM is implemented as an add-on package to MALLET. It makes heavy use of MALLET's data structures and optimization facilities. Its name stands for GRaphical Models in Mallet. It has been developed for several years, and results from GRMM have been used in several papers. It is written by Charles Sutton.
Sutton, Charles. "GRMM: GRaphical Models in Mallet." http://mallet.cs.umass.edu/grmm/. 2006.