SEER：基于选择性视觉-文本压缩的长上下文推理
SEER: Long-Context Reasoning via Selective Visual-Text Compression
- The University of North Carolina at Chapel Hill（北卡罗来纳大学教堂山分校）
- University of Cambridge（剑桥大学）
- University of Technology Sydney（悉尼科技大学）
- The University of North Carolina at Charlotte（北卡罗来纳大学夏洛特分校）
- The University of Texas at Austin（德克萨斯大学奥斯汀分校）
机构由 AI 辅助整理，请以论文原文为准。
中文总结 AI 辅助
SEER是结合视觉压缩效率与文本推理精度的框架，经监督微调后在LongBench等长上下文基准测试中，准确率优于Glyph-9B、Qwen3-8B等基线模型，可提升提取精度并保留提示token节省量。
中文摘要 AI 辅助
对于大语言模型而言，由于文本token注意力的二次复杂度，长上下文推理的计算成本仍然很高。视觉-文本压缩通过将文本渲染为图像并使用视觉-语言模型处理，通常可减少token使用量，是一种很有前景的替代方案。然而，现有方法会应用均匀压缩，不考虑查询相关性，在需要详细提取的地方可能会牺牲精度。我们提出SEER，这是一个通过视觉扫描学习选择与查询相关的图像、仅在需要时检索文本内容的框架，结合了视觉压缩的效率和基于文本推理的精度。通过对工具交互轨迹进行监督微调，SEER学习了用于选择和检索的自适应工具调用。在长上下文基准测试中，SEER通过选择性文本检索提高了提取精度，同时保留了相对于全文基线的平均提示token节省量。在LongBench上，SEER实现了51.11%的平均准确率，超过视觉-文本基线Glyph-9B 2.33个百分点，超过Qwen3-8B 3.49个百分点。代码可在此URL访问
英文摘要
Long-context reasoning remains computationally expensive for large language models due to the quadratic complexity of attention over text tokens. Visual-text compression offers a promising alternative by rendering text into images and processing them with vision-language models, often reducing token usage. However, existing approaches apply uniform compression regardless of query relevance, potentially sacrificing precision where detailed extraction is required. We present SEER, a framework that learns to select query-relevant images through visual scanning and retrieve textual content only where needed, combining the efficiency of visual compression with the precision of text-based reasoning. Through supervised fine-tuning on tool-interaction trajectories, SEER learns adaptive tool invocation for selection and retrieval. Experiments on long-context benchmarks show that SEER improves extraction precision through selective text retrieval while retaining average prompt-token savings relative to full-text baselines. On LongBench, SEER achieves 51.11% average accuracy, outperforming the visual-text baseline Glyph-9B by 2.33 points and Qwen3-8B by 3.49 points. Code can be accessed at https://github.com/jiaweixu98/SEER
