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

Summary

This paper introduces a self-evolving training framework that enables unified large multimodal models (LMMs) to improve both visual understanding and image generation autonomously, without curated post-training supervision such as human annotations, preference labels, or external reward models. The approach assigns three internal roles within a single model: a Proposer that generates visual questions, a Solver that answers and evaluates them, and a Generator that produces images based on those questions. Training is driven by self-consistency rewards computed only from unlabeled images, allowing the model to iteratively refine its understanding and generation capabilities in a fully self-supervised loop. The work addresses the dependence of most unified LMMs on curated supervision and demonstrates a path toward continuous self-improvement using unlabeled image data alone. Authored by Ritesh Thawkar, Shravan Venkatraman, and Omkar Thawakar, the paper (arXiv:2606.27376) falls in the computer vision research area.