🧪 Test?View on arXiv
Mage-VL-Cowork: A Full Training Recipe to Learn Hybrid Computer Use From Evolved Harnesses
Not provided in the content
multimodalreinforcement learningcurriculum learningagent training
2610.40747
Builder Relevance
1h ago80%
Abstract
Mage-VL-Cowork presents a hybrid computer-use agent trained through a progressive curriculum, achieving state-of-the-art performance in multimodal interaction.
Reality Card
Core Claim
Mage-VL-Cowork-4B achieves 66.0% on OSWorld-Verified, demonstrating state-of-the-art performance at its parameter size through an efficient training curriculum.
Method / Result
Reduced recurring per-task expenditure by 82.6% while synthesizing 293,806 tasks across 6,915 skills.
Limitations
The paper does not provide specific details on the reproducibility of the training process or the generalization of the model across different environments.
Paper to code
Verified implementation resources so builders can test the paper’s claims instead of stopping at the abstract.
No verified implementation link has been attached yet. AIBuzzHub will keep this panel separate from unverified search results.