Ask, Solve, Generate: Self-Evolving Unified Multimodal Understanding and Generation via Self-Consistency Rewards

arXiv·medium signal

Proposes a self-evolving unified large multimodal model that improves both visual understanding and image generation using self-consistency rewards instead of curated human annotations or preference labels. Removes the post-training supervision bottleneck that constrains most unified LMMs. Of interest to teams training multimodal models without large annotation budgets.