💡 This paper proposes ARM, a discrete representation-based autoregressive large multimodal model that unifies image understanding, generation and editing via multi-objective supervised visual tokenizer