IMUG-Bench: Benchmarking Unified Multimodal Models on Interleaved Understanding and Generation

Researchers have introduced new unified multimodal models (UMMs) that support both understanding and generation within a single framework, addressing limitations in existing benchmarks for evaluating multi-turn interleaved image-text dialogues.

Unified multimodal models have emerged to handle both understanding and generation tasks, according to a paper submitted to arXiv^[1]. However, existing benchmarks fail to assess multi-turn interleaved image-text dialogues, a crucial task for UMMs in real-world applications. To address this gap, researchers proposed IMUG-Bench, a comprehensive benchmark that evaluates UMMs' understanding and generation capabilities across three classes: Static Spatial, Temporal Causal, and Hybrid. IMUG-Bench covers 3,113 samples and 12,034 interaction turns^[1]. Separately, a team introduced HYDRA-X, a unified multimodal model that unifies image and video tokenization within a single Vision Transformer, addressing challenges such as injecting spatiotemporal reconstruction capability and embedding semantic awareness into the latent space. Comprehensive ablations showed that frame-level causal temporal attention suffices for visual reconstruction, while full spatiotemporal attention degrades it^[2].