Consistency Should Be the Priority for Unified Multimodal Models

Unified Multimodal Models (UMMs) emerge as a promising paradigm for integrating understanding and generation within a single architecture. This paper argues that current approaches achieve only interface-level unification where inconsistencies are the primary bottleneck. We identify critical dimensions where consistency must be established: (1) Consistency-Aware Asymmetric Grounding, where models must account for uneven or incomplete evidence reliability rather than treating all modalities as equally informative; (2) Consistency-Aware Representation Learning, where the representation trade-off between understanding and generation should be modeled; (3) Consistency-Aware Reasoning, where multi-step and cross-task consistency should be explicitly modeled; (4) Consistency-Aware Post-Training, where specific training should be considered for the conflicting objective for understanding and generation pathways; and (5) Consistency-Aware Evaluation, where benchmarks must shift from measuring isolated capabilities to assessing consistency. We contend that the next leap in unified multimodal intelligence requires resolving these systemic inconsistencies.