Prior works on Information Extraction (IE) typically predict different tasks and instances (e.g., event triggers, entities, roles, relations) independently, while neglecting their interactions and leading to model inefficiency. In this work, we introduce a joint IE framework, HighIE, that learns and predicts multiple IE tasks by integrating high-order cross-task and cross-instance dependencies. Specifically, we design two categories of high-order factors: homogeneous factors and heterogeneous factors. Then, these factors are utilized to jointly predict labels of all instances. To address the intractability problem of exact high-order inference, we incorporate a high-order neural decoder that is unfolded from a mean-field variational inference method. The experimental results show that our approach achieves consistent improvements on three IE tasks compared with our baseline and prior work.