Trident: New Reranker Enhances Long-Document Multimodal QA
A recent paper on arXiv (2608.14841) presents Trident, an innovative system aimed at enhancing reranking in long-document visual question answering (VQA). It identifies a significant issue in retrieve-then-read frameworks: although retrieval recall is commendable, the selection of evidence on the reranker side is inadequate. For instance, BGE-M3 scores a Recall@20 of 0.86 but only an F1@5 of 0.254 on the MMLongBench-Doc benchmark, while the visual retriever ColPali achieves an F1@5 of 0.332. A text-only rerank LLM, which relies solely on raw snippets, overlooks crucial evidence like tables and charts. Trident consists of two parts: Trident-R, a retriever-agnostic LLM reranker, and Trident-S, a generation-side module. The paper is a new submission and can be accessed via the provided URL.
Key facts
- Paper arXiv:2608.14841 introduces Trident for long-document multimodal QA.
- Trident addresses reranker-side evidence selection bottleneck.
- On MMLongBench-Doc, BGE-M3 achieves Recall@20 = 0.86 but F1@5 = 0.254.
- ColPali visual retriever reaches F1@5 = 0.332.
- Text-only rerank LLM misses table, chart, and layout evidence.
- Trident-R converts candidates into semantic records with visual caption, section path, entity tags, multi-axis concept hits, and text snippet.
- Trident-R performs a single adaptive-K rerank call.
- Trident-S prompts the VLM under topical, entity, and other constraints.
Entities
Institutions
- arXiv