Tencent Hunyuan Joins Universities to Release OpenSearch-VL Open-Source Multimodal Search Training Scheme

Tencent Hunyuan recently collaborated with the University of California, Los Angeles (UCLA), the Chinese University of Hong Kong, and other institutions to release the OpenSearch-VL open-source multimodal training scheme. Using reinforcement learning (RL) technology, this scheme builds deep search agents that can process multi-modal inputs like images and text, and actively call external tools for reasoning and retrieval, aiming to solve complex visual question-answering problems.

The research team pointed out that the current bottleneck in the development of multimodal search agents lies in high-quality training data. OpenSearch-VL provides a complete open-source solution, including data pipelines, tools, and training algorithms. Its data pipeline uses Wikipedia hyperlink sampling to generate the SearchVL-SFT-36k dataset; the tool environment integrates text search, image processing, and other functions. Experiments show that the relevant model score increased from 47.8 to 61.6, with significant results.