Junke Wang, Xiao Wang, Jiacheng Pan, Xuefeng Hu, Feng Li, Jingxiang Sun, Chaorui Deng, Zilong Chen et al.
ARM is a large multimodal model that combines a discrete semantic visual tokenizer with an autoregressive model to unify image understanding, generation, and editing within a single next-token prediction framework.
Existing multimodal models use different architectures or representations for image understanding and generation, making unification difficult, and they lack preference alignment, especially in generation tasks.
1) Train a discrete semantic visual tokenizer that jointly optimizes semantic discriminability, language alignment, and faithful reconstruction. 2) Train a 7B autoregressive model on large-scale text and image token sequences to develop vision-language perception and generation capabilities. 3) Apply reinforcement learning (RL) to optimize task-level objectives such as visual quality, instruction adherence, and edit consistency.
After applying RL, the overall WISE score improved from 0.50 to 0.56, and GEdit-Bench-EN G_O from 5.75 to 6.68. Cross-task synergy between text-to-image generation and editing was observed, demonstrating that autoregressive modeling combined with strong representations and preference optimization can serve as a scalable foundation for multimodal intelligence.