📄 Paper of the Day2026-07-12
Representation Forcing for Bottleneck-Free Unified Multimodal Models2026 · Vision-Language Models
Unified multimodal models (UMMs) aim to handle perception and generation in a single model. Yet existing UMMs still rely on a frozen, separately pretrained VAE for image generation, imposing a structural bottleneck. Naiv…