Trident Improves Long-Document VQA with Multi-Aspect Annotation.

This research addresses the bottleneck in long-document visual question answering (VQA), where current retrieve-then-read pipelines struggle with reranking and evidence selection, especially when documents contain mixed modalities like text, tables, and images. Even advanced retrievers often fail to provide sufficient context for text-only rerankers, leading to poor performance. The proposed solution, Trident, aims to improve this by enriching the information available to the reranker and the VLM during generation. Trident consists of two main components. Trident-R is a retriever-agnostic LLM reranker that transforms each retrieved document candidate into a comprehensive semantic record. This record includes not only text snippets but also visual captions, section paths, entity tags, and multi-axis concept hits, allowing the LLM to make more informed reranking decisions. Trident-S, the second component, is a generation-side module that guides the VLM by prompting it with topical, entity, and structural lenses before synthesizing the final answer. The results show that this annotation-plus-rerank protocol substantially boosts retrieval F1 scores across various document pools and significantly improves generation accuracy for open-ended synthesis questions, demonstrating the critical role of structured annotation in multimodal long-document VQA.