Researchers have developed a new system called Trident for answering questions about long documents that contain text, tables, charts, and figures. The system consists of two components: Trident-R, which converts candidate answers into a format that can be understood by a language model, and Trident-S, which generates a visual caption and other information to help the language model answer questions accurately. In experiments, Trident outperformed existing systems in retrieving relevant information from long documents, with improvements in accuracy ranging from 6.6 to 12.4 points. The authors attribute the success of Trident to its ability to provide structured annotation to the language model, which helps it focus on relevant information. --- Why it matters: This matters because current question-answering systems often struggle to accurately retrieve and understand information from long documents that contain multiple types of content. Trident's improvements in accuracy can help improve the performance of downstream applications such as search engines, knowledge graphs, and expert systems. Source: https://arxiv.org/abs/2608.14841