SAIL-Unify: Hierarchical Interleaving Augmented Holistic Bilateral Synergy in Unified Multimodal Models
Abstract
Human cognition is fundamentally driven by a bilateral synergy loop, where multimodal understanding (MMU) and generation interleave and reinforce each other. This observation motivates Unified Multimodal Models (UMMs) to foster bilateral synergy across multiple dimensions, such as data and architecture. However, existing methods typically explore these dimensions in isolation and remain confined to limited training stages, leaving the hierarchical role of interleaving in advancing holistic bilateral synergy largely underexplored. To this end, we propose SAIL-Unify, a new UMM that systematically explores the synergistic gains of hierarchical interleaving across data, architectural design, learning strategies, and shared policy optimization throughout the full lifecycle of pre-training and post-training. Specifically, we first unify data formats and establish dual-track pipelines for rigorous filtering and new data construction, yielding over 179 million samples. Architecturally, we introduce the Cross-Level Adaptive Modulation (CLAM) Component within the native UMM paradigm to better harmonize bilateral representational flows. To stabilize task synergy, we further propose a curriculum-like data arrangement strategy coupled with a progressive training recipe, aligning task difficulty with model capacity for robust convergence. Moreover, we develop a new shared policy optimization framework established through a reciprocal causal loop, which mitigates gradient mismatch in the unified RL space and more fully unlocks the potential of pre-training. Extensive quantitative and qualitative experiments across nine benchmarks demonstrate the effectiveness of proposed SAIL-Unify in advancing both MMU and visual generation. Ablation studies further shed light on the crucial factors behind its performance.