Overview
Research Area: Computer Vision (CV) Authors: Mingfei Gao, Rui Tian, Haiming Gang, Bohan Zhai, Le Zhang, Yuanzheng Gong, Di Feng, Ege Özsoy, Kaixin Ma, Vishwesh Kirthivasan, Oğuzhan Fatih Kar, Roman Bachmann, Anders Boesen Lindbo Larsen, Afshin Dehghan Published: 2026-09-18 arXiv: 2609.22083
Summary
The authors present MintAct, a family of vision-language models that unifies UI grounding, multi-step navigation across mobile, desktop, and web, and visual tool use, trained at 2B, 4B, and 8B scales. Through careful design of environments, data, and training recipes, MintAct models match the performance of per-domain specialists across all of these capabilities.
To enable this, the team develops a scalable environment and reinforcement learning (RL) infrastructure. On the environment side, they host hundreds of concurrent instances across heterogeneous per-domain backends, serving both trajectory data collection and online RL. To enable efficient and scalable RL training, an asynchronous framework keeps explicit control over the cross-domain training distribution and remains stable under noisy environment feedback and off-policy drift.
Experiments show that MintAct achieves state-of-the-art performance among models of the same size on multiple benchmarks, including OSWorld-Verified 48.9.
Key Points
- Unified family of vision-language models covering UI grounding, multi-step navigation (mobile, desktop, web), and visual tool use
- Trained at 2B, 4B, and 8B scales; matches per-domain specialist models
- Scalable RL infrastructure with hundreds of concurrent environment instances across heterogeneous backends
- Asynchronous training framework with explicit control over cross-domain data distribution, stable under noisy feedback and off-policy drift
- State-of-the-art at equal model size: OSWorld-Verified 48.9
*Auto-collected on 2026-09-22*