BEIJING, Oct. 6, 2025 — In the world of artificial intelligence, the term “Agent” is gaining a lot of attention. Experts believe that for Agents to be truly helpful, they need to learn how to use mobile phones and computers just like humans do, interacting with visual interfaces.
MiningLamp Technology, a top Chinese company in advanced data intelligence and large models, has announced that its specialized GUI model, Mano, has achieved remarkable success in two important benchmark tests: Mind2Web and OSWorld. Thanks to two innovative techniques—online reinforcement learning and automated data gathering—Mano represents a new way for developing GUI agents that can learn and improve over time.
Key Achievements:
-
Mind2Web: Mind2Web tests Agents on 137 websites and over 2,350 real-life tasks, ensuring they can accurately identify targets in complex structures and perform specific actions. Mano excelled here, showcasing its ability to “see clearly and deliver results.” Reports indicate that Mano outperformed others in both Element Accuracy and Step Success Rate, establishing new standards for precision in task execution.
-
OSWorld-Verified: This test posed a tougher challenge, featuring 369 tasks across various applications. Mano achieved a success rate of 41.6%, surpassing other models such as Qwen and GUI-Owl.
Technical Innovations:
Highlight One: Online Reinforcement Learning
Traditionally, many models trained with offline reinforcement learning rely on pre-collected data. However, Mano introduces the concept of “online reinforcement learning,” enabling continuous learning from real-time interactions. This method helps maintain a balance between trying new actions and executing known strategies.
MiningLamp Technology also created a simulation environment for Agents to gather diverse data through real-world interests. This approach overcomes limitations of offline learning, resulting in improved performance in various web GUI scenarios. By combining online sampling with offline filtering, Mano effectively addresses learning inefficiencies.
Initial studies revealed that after implementing online reinforcement learning, Mano’s average score significantly increased, reaching a total of 41.6.
Highlight Two: Intelligent Data Collection
While large models can often handle general instructions, they can struggle with detailed tasks that require multiple steps. Historically, collecting high-quality training data has been time-consuming. To fix this, MiningLamp Technology created a system that automates data collection, greatly enhancing the efficiency and quality of data acquired.
Mano operates within a virtual environment that simulates interactions, generating realistic data for training. It uses a specialized Chrome plugin to extract relevant elements from web environments and applies advanced techniques for desktop interactions.
To further boost data collection intelligence, the team developed a smart exploration tool that selects interactive elements effectively. This ensures the captured data is relevant and high-quality, allowing Mano to learn better.
Mano’s remarkable performance is a result of MiningLamp Technology’s years of expertise in building large models. In 2024, they made great strides in processing non-standard data with their models. With the introduction of DeepMiner, a reliable intelligent agent for data analysis, Mano is set to take intelligent operations in complex software environments to the next level. Moving forward, MiningLamp aims to enhance Mano’s capabilities further for real-world applications, helping businesses evolve intelligently.
Source: MiningLamp Technology
