Papers & Research

Answer.AI Releases ReadBench, Revealing Significant Performance Degradation of VLMs on Long-Text Image Reading

Original title:TIL: Vision-Language Models Read Worse (or Better) Than You Think

AI summary

Answer.AI launched ReadBench, a new benchmark specifically evaluating the ability of vision-language models (VLMs) to extract and reason over textual information from images. The study found that while high resolution has little impact on generation tasks, nearly all VLMs show significant performance degradation when processing multi-page long-document images, indicating that current Visual RAG approaches are not yet viable for long contexts. By comparison, GPT-4o performed relatively well in short-context and multi-page scenarios, with smaller overall degradation. The benchmark is open-source and covers image versions of datasets including MMLU and GPQA.

Key Takeaways

  • Answer.
  • AI launched ReadBench,a new benchmark specifically evaluating the ability of vision-language…
  • The study found that while high resolution has little impact on generation tasks,nearly all VLMs…
  • By comparison,GPT-4o performed relatively well in short-context and multi-page scenarios,with…
  • The benchmark is open-source and covers image versions of datasets including MMLU and GPQA.

Quick Facts

Published
Updated
Source
Answer.AI 官方研发博客(RSS)
Category
Papers & Research
AI score
78
Read the original Read the full article at the original source
Why it matters

针对热门的多模态检索增强生成(Visual RAG)场景提供了关键实证数据,揭示了 VLM 在长文档阅读上的实际瓶颈,为开发者选择模型和优化管道提供重要参考。

Source & copyright

This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.

Original article · AIHOT entry

How to Cite

When citing this article, keep the title, date, and permalink, and credit Orange News as the source.

Orange News: Answer.AI Releases ReadBench, Revealing Significant Performance Degradation of VLMs on Long-Text Image Reading. June 5, 2025. https://new.nat6.net/en/news/nqiq403r0cs6cekuujaxe5n7b/

View all →