Mixture of Contexts (MoC) was proposed by ByteDance Seed team in collaboration with Stanford University and other institutions in August 2025. The relevant research results were published in the paper "Mixture of Contexts for Long Video Generation".

Mixture of Contexts (MoC) is a simple, learnable sparse attention routing module based on reframing long-context video generation as an internal information retrieval task. In MoC, each query dynamically selects a few informative segments and mandatory anchors for attention, preventing loop closure through causal routing.