ARM: An Autoregressive Large Multimodal Model With Unified Discrete Representations

arXiv (via HuggingFace Daily Papers)·low signal

ARM (arXiv 2606.11188) unifies text, image, and audio under shared discrete token representations in a single autoregressive multimodal model, sidestepping separate modality-specific encoders. For builders it's another data point in the shift toward one tokenizer/one model handling all modalities natively rather than bolting on adapters.