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

Summary

This paper introduces a self-evolving training framework for unified large multimodal models (LMMs) that support both visual understanding and image generation. Most existing unified LMMs rely on curated post-training supervision such as human annotations, preference labels, or external reward models. The proposed approach enables a model to improve autonomously using only unlabeled images through three internal roles: a Proposer that generates visual questions, a Solver that answers and evaluates them, and a Generator that produces images conditioned on those questions. Training is driven by self-consistency rewards, allowing the model to iteratively refine both its visual understanding and generation capabilities without external supervision. The work was posted to arXiv under identifier 2606.27376 by Ritesh Thawkar, Shravan Venkatraman, and Omkar Thawakar in the computer vision domain.