In this paper, we explore how LFG analyses as produced by the XLE parser with the English ParGram grammar can be used in a probabilistic coreference resolution system. So far, such systems have mainly relied only on information from surface-based NLP tools, reaching reasonable levels of performance while requiring only small amounts of training data. We compare these surface-based approaches with a first attempt at an LFG-based coreference system and another system using the treebank-trained probabilistic parser by Charniak. Based on the (limited) quantity of training data we used, the performance of all three approaches was quite comparable. However, there are some indications that an XLE-based approach may lead to better results if trained on larger training sets.