Marvo

Marvo is a visual Agent training method that self-evolves from unlabeled images, enabling the Agent to learn to effectively invoke tools without any human supervision.

Marvo adopts a dual-Agent architecture consisting of a Challenger and a Solver: the Challenger automatically generates question-answer pairs from unlabeled images, and the Solver answers these questions through reinforcement learning.

Marvo's Margin Reward is defined as the difference in the Solver's performance between using tools and not using tools.

The research team validated Marvo's effectiveness on multiple benchmarks, covering scenarios such as Agent reasoning, visual search, mathematical reasoning, and general real-world reasoning, with results showing an average accuracy improvement of up to 7.1%.

Marvo's improvement does not come from more annotated data or a larger model, but from the design of the training signal.

Marvo is suitable for multimodal scenarios where data acquisition is costly, such as medical imaging or remote sensing imagery that require specialized annotation.

The stability of self-evolution methods, the handling of noise in reward signals, and cross-modal transfer capability remain questions that subsequent research needs to answer.

Marvo's Challenger does not seek to defeat the Solver, but to generate questions that precisely require tools, so it is less competitive and more collaborative.

Marvo's design reduces the risks of mode collapse or uncontrolled difficulty that easily arise in self-play, making training more stable and controllable.