Unified multimodal models (UMMs) embed both understanding and generation in a single backbone, yet these capabilities often clash, raising the question of how representations and computation should be shared across functions. Most existing UMMs rely on a largely fixed architecture where all inputs and functionalities follow the same depth and parameter‑sharing pattern. We argue that this assumption is overly restrictive: effective unification requires shared computation for cross‑functional communication while preserving task‑ or input‑specific capacity. To address this, we introduce FlexUMM, a flexible UMM framework that dynamically decides when to keep shared computation, when to route specialized capacity, and when to halt further processing. Empirical results show that this flexible allocation improves UMMs along two complementary axes: it boosts accuracy by reducing harmful competition between functions and enhances efficiency by avoiding unnecessary computation once an input representation is mature enough. Beyond simple architectural unification of pre‑trained models, FlexUMM serves as a model‑agnostic functional plug‑in that makes existing UMMs more unified, stronger, and more efficient. Code is available at https://github.com/AIFrontierLab/TorchUMM/tree/main/src/umm/post_training/flexumm
Review