LVLM Abilities for Long-Context Document Understanding

Establish the capabilities of Large Vision-Language Models (LVLMs) for long-context document understanding by determining whether these models can reliably understand and answer questions over lengthy, multi-page documents.

Background

Large Vision-LLMs have achieved strong results on single-page document understanding benchmarks, yet real-world documents often span tens of pages and contain diverse modalities and complex layouts. This creates additional challenges beyond single-page tasks, such as localization across large contexts and cross-page reasoning.

The paper introduces MMLongBench-Doc precisely to evaluate this gap and reports that current LVLMs struggle significantly on long-context document understanding, underscoring the need to rigorously determine and characterize LVLM capabilities in this setting.

References

However, their abilities on long-context DU remain an open problem.

— MMLongBench-Doc: Benchmarking Long-context Document Understanding with Visualizations  (2407.01523 - Ma et al., 2024) in Abstract, page 1

This leaves the cost/accuracy trade-off across VLM scale largely open.

— An Empirical Study of VLM Pipelines for Long-Document QA  (2609.29933 - Ak et al., 24 Sep 2026) in Related Work, Section 2.2, subsection “Vision-Language Models”

Finally, our results assume English-language documents and the VLM families we tested, and whether the rankings carry over to other languages and to smaller on-device VLMs is open.

— An Empirical Study of VLM Pipelines for Long-Document QA  (2609.29933 - Ak et al., 24 Sep 2026) in Limitations