Crowdworks said on Monday it has won a contract to build “complex document-based knowledge inference data” under the Ministry of Science and ICT-led 2026 AI Hub training data (inference) programme promoted by the National Information Society Agency (NIA).

The project will be carried out by a consortium led by Crowdworks with Wart Intelligence and Zendive participating.

Recent Korean-language specialist knowledge evaluations of generative AI at home and abroad showed the inference capabilities of many AI models fell short of the human average. Limits in inference capabilities, such as understanding context from multiple angles and deriving causal relationships, were cited as the cause.

The ministry planned the AI Hub training data (inference) programme to address such limits. Out of a total of 10 project tasks, the Crowdworks consortium will take charge of building “complex document-based knowledge inference data”.

The consortium will collect and refine multimodal public documents that combine visual elements such as tables, graphs and diagrams with text, and produce more than 10,000 cases of high-quality inference training data. It will build an original dataset for Korean multimodal complex inference.

Crowdworks will analyse weak inference areas by field, including public administration, science and technology, and laws and patents. It will design complex chains of thought and high-difficulty questions using inference techniques such as chain-of-thought (CoT), tree-of-thought (ToT) and graph-of-thought (GoT).

To prevent hallucinations, the company will match sources at each step of inference and apply a “hybrid verification pipeline” combining large language model (LLM) automated verification with reviews by expert workers to ensure data integrity, Crowdworks said.

A Crowdworks official said completing the project would contribute to securing proprietary AI technology and data sovereignty that understand Korea-language-based administrative and industrial contexts. The official added the company would contribute to strengthening the competitiveness of the domestic AI industry with its accumulated know-how in data design and processing and its quality management capabilities.