Agent0-VL: Exploring Self-Evolving Agent for Tool-Integrated Vision-Language Reasoning

Large Vision-Language Models (LVLMs) have achieved remarkable progress in multimodal reasoning tasks; however, their learning remains constrained by the limitations of human-annotated supervision. Recent self-rewarding approaches attempt to overcome this constraint by allowing models to act as their own critics or reward providers. Yet, purely text-based self-evaluation struggles to verify complex visual reasoning steps and often suffers from evaluation hallucinations. To address these challenges, inspired by recent advances in tool-integrated reasoning, we propose Agent0-VL, a self-evolving vision-language agent that achieves continual improvement with tool-integrated reasoning. Agent0-VL incorporates tool usage not only into reasoning but also into self-evaluation and self-repair, enabling the model to introspect, verify, and refine its reasoning through evidence-grounded analysis. It unifies two synergistic roles within a single LVLM: a Solver that performs multi-turn tool-integrated reasoning, and a Verifier that generates structured feedback and fine-grained self-rewards through tool-grounded critique. These roles interact through a Self-Evolving Reasoning Cycle, where tool-based verification and reinforcement learning jointly align the reasoning and evaluation distributions for stable self-improvement. Through this zero-external-reward evolution, Agent0-VL aligns its reasoning and verification behaviors without any human annotation or external reward models, achieving continual self-improvement. Experiments on chart reasoning, geometric problem solving, and visual scientific analysis show that Agent0-VL achieves an 12.5% improvement over the Qwen-VL base model.

AI models that understand both images and text, known as vision-language models, are typically improved using human feedback, a process that is expensive and hard to scale. A newer approach lets models evaluate their own work, but when tasks involve visual reasoning, such as reading charts or solving geometry problems, text-only self-assessment often fails: the model might "think" its answer looks right without actually verifying the math or checking what the image shows. Agent0-VL solves this by giving the model two alternating roles within a single system. As a Solver, it reasons step by step and calls external tools like a code interpreter to perform precise calculations. As a Verifier, it reviews each step, re-running tools to fact-check results against ground truth. When the Verifier catches an error, a self-repair mechanism patches the faulty step and lets the Solver try again. This solve-verify-repair cycle repeats across multiple training rounds, allowing the model to continuously improve without any human grading. On seven visual reasoning benchmarks spanning math, chart analysis, and hallucination detection, Agent0-VL improves accuracy by 12.5% over its base model and outperforms all existing open-source systems of comparable size.