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 depend on curated post-training supervision such as human annotations, preference labels, or external reward models. The authors propose an alternative: a framework with three internal roles — a Proposer that generates visual questions from unlabeled images, a Solver that answers and evaluates those questions, and a Generator that produces images conditioned on the questions. Using self-consistency rewards, the model iteratively improves both its understanding and generation capabilities without external supervision. Experiments reported in the paper show significant improvements on both understanding and generation tasks, demonstrating that self-evolution is a viable path for training unified multimodal models. The work was posted on arXiv (2606.27376) by Ritesh Thawkar, Shravan Venkatraman, and Omkar Thawkar, and falls within the computer vision research area. It is relevant to researchers interested in self-supervised learning, multimodal model training, and reward-model-free post-training methods.