Chinese artificial intelligence (AI) researchers have released an open-source framework and a real-world scenario competition platform that significantly improves AI industrial application.
As AI computing power grows stronger and large language models become increasingly sophisticated, Chinese researchers have been focused on how best to apply it to real-world scenarios.
To this end, a research initiative at the Artificial Intelligence Innovation Center at Yangtze Delta Region Institute of Tsinghua University standardized human-machine interaction, task-set mechanisms and human feedback systems, resulting in enhanced industrial application efficiency and greater enterprise deployment.
The team leader said that the global AI industry currently faces a structural contradiction: the exponential growth of model and tool capabilities versus the linear climb in industrial adoption rates. The core contradiction in AI development has shifted from "enhancing model intelligence" to "bridging the deployment gap."
REAL-WORLD FRAMEWORK
To address the gap between AI capabilities and real-world deployment, the team released the Real World AI (RWAI) open-source framework, expanding the scope of open-source efforts from code and tools to encompass role definitions, workflow design, human-machine interaction, and human-human collaboration as an integrated practice.
The framework reconstructs the interaction between AI and humans in real-world tasks through three core elements: restoring real-world task sets, capturing authentic human feedback from real interactions, and standardizing human-machine interaction protocols.
According to the team, the real-world tests have proven that RWAI outperforms traditional software development models in practical efficiency, actual effectiveness and resolution time, reducing pre-project validation timelines from two-to-three months to less than two weeks.
The team also launched the AI arena platform. Unlike traditional benchmarks or model leader boards, the platform focuses on evaluating the actual effectiveness of AI solutions in real business operations, including metrics such as organizational costs, time efficiency, computing costs and compliance requirements.
The platform adopts a "challenger-champion" mechanism, where competing entities are not single models but complete solutions, encompassing team configurations, workflows, agent combinations, and context engineering. The best-practice workflows corresponding to winning solutions will be made public and available for replication.
INDUSTRIAL APPLICATIONS
China Southern Power Grid's internet service subsidiary utilized RWAI platform to address the end-to-end safety management challenges of power grid infrastructure projects, ranging from planning to on-site execution. Faced with complex compliance requirements, traditional manual supervision on infrastructure projects had reached an efficiency bottleneck.
Using the platform, the company developed an intelligent risk control solution for on-site and subcontractor management, increasing hidden risk detection rates by approximately 40 percent and boosting risk warning accuracy to 92 percent. The company and research team are now preparing to advance a demonstration project for generative AI across the full lifecycle of construction planning.
The company's senior engineer Hu Rui said the RWAI platform had successfully bridged the gap between AI technology and deployment, while significantly reducing trial-and-error costs. The system has transformed engineering management from reactive response to proactive intelligent control, using AI to support high-quality power grid construction.
Jiangsu Eastern Shenghong Co., Ltd. also used the RWAI platform. As a petrochemical manufacturer, it has long faced challenges such as integrating knowledge in traditional process industries, applying general-purpose AI to core business operations, and the lack of controllability in decision-making by large models.
Leveraging the RWAI platform, Eastern Shenghong integrated 30 years of production process knowledge with data from the full industry chain, overcoming the scarcity of chemical industry corpora and high-compliance barriers to build an industrial large model that truly understands the business.
Through multimodal fault monitoring and prediction, the company has significantly reduced unplanned downtime in key production lines and can dynamically recommend optimal production scheduling, achieving cost reduction, efficiency improvement, and process optimization.
Yang Tianwei, vice chairman of the company and general manager of its AI business unit, said that by using the RWAI platform's evaluation capabilities, they have transformed internally validated, high-quality model capabilities into a library of reusable, billable, and composable products.
"This not only activates Eastern Shenghong's own intelligent development, but also provides a field-proven and best-practice solution for deploying large models in the process industries," Yang added.
The RWAI platform now covers multiple application scenarios including industrial forecasting systems, document review and risk control, and research report generation. Its implementations have already been deployed in projects for some Fortune Global 500 companies.
The research team said that the platform will also supply real-world human-computer interaction data to support large model development and academic research.
Reviewer 1: Huang Mengyao
Reviewer 2: Zhang Yanlan
Reviewer 3: Tang Caihong