Ask, Solve, Generate: Self-Evolving Unified Multimodal Models via Self-Consistency Rewards

Summary

This paper introduces a self-evolving training framework for unified large multimodal models (LMMs) that handle 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 proposed approach instead enables a model to improve autonomously using only unlabeled images, without external supervision. The framework assigns three internal roles to the model: a Proposer that generates visual questions, a Solver that answers and evaluates those questions, and a Generator that produces images based on the questions. Training is driven by self-consistency rewards, allowing the model to iteratively enhance both its understanding and generation capabilities. The work is in computer vision, authored by Ritesh Thawkar, Shravan Venkatraman, and Omkar Thawakar, and available on arXiv (2606.27376).