Paper 2603.09206
MM-Zero: Self-Evolving Multi-Model Vision Language Models From Zero Data
01 In brief
Summary
MM-Zero is a reinforcement learning framework that enables zero-data self-evolution for Vision Language Models (VLMs).
It introduces a tri-role system—Proposer, Coder, and Solver—all initialized from the same base model and trained sequentially using Group Relative Policy Optimization (GRPO).
The Proposer generates visual descriptions and questions, the Coder translates them into executable code (e.g., SVG) to render images, and the Solver performs multimodal reasoning on the generated visuals.
Reward mechanisms include execution feedback, solvability scoring, difficulty balancing, and diversity penalties.
Experiments on Qwen3-VL-4B/8B and Mimo-VL-7B show consistent improvements across benchmarks like MMMU, ChartQA, MathVista, and HallusionBench, with the 8B model improving from 50.7% to 54.1% average accuracy.
Ablations show that removing solvability capping leads to reward hacking, and removing content diversity causes overfitting.
The framework demonstrates scalable self-improvement without external data, though scaling to larger models remains untested due to computational costs.
02 From the paper
Abstract
Self-evolving has emerged as a key paradigm for improving foundational models such as Large Language Models (LLMs) and Vision Language Models (VLMs) with minimal human intervention. While recent approaches have demonstrated that LLM agents can self-evolve from scratch with little to no data, VLMs introduce an additional visual modality that typically requires at least some seed data, such as images, to bootstrap the self-evolution process. In this work, we present Multi-model Multimodal Zero (MM-Zero), the first RL-based framework to achieve zero-data self-evolution for VLM reasoning. Moving beyond prior dual-role (Proposer and Solver) setups, MM-Zero introduces a multi-role self-evolving training framework comprising three specialized roles: a Proposer that generates abstract visual concepts and formulates questions; a Coder that translates these concepts into executable code (e.g., Python, SVG) to render visual images; and a Solver that performs multimodal reasoning over the generated visual content. All three roles are initialized from the same base model and trained using Group Relative Policy Optimization (GRPO), with carefully designed reward mechanisms that integrate execution feedback, visual verification, and difficulty balancing. Our experiments show that MM-Zero improves VLM reasoning performance across a wide range of multimodal benchmarks. MM-Zero establishes a scalable path toward self-evolving multi-model systems for multimodal models, extending the frontier of self-improvement beyond the conventional two-model paradigm.