Research Papers 论文研究 3h ago Updated 1h ago 更新于 1小时前 49

Infinity-Parser2 Technical Report Infinity-Parser2 技术报告

Infinity-Parser2 introduces a novel architecture combining a controllable data-synthesis engine with multi-task reinforcement learning to solve the scarcity of high-quality document parsing data. The team released Infinity-Doc2-5M, a 5-million-sample bilingual corpus featuring precise annotations for bounding boxes, reading order, and various content formats like Markdown, LaTeX, and chemical structures. A verifiable, multi-task reward system enables Joint Reinforcement Learning across eight dis Infinity-Parser2 提出了一种结合可控数据合成管道与多任务强化学习的端到端文档解析大模型方案。 构建了开源的 Infinity-Doc2-5M 数据集,包含500万双语样本,涵盖多种文档类型及结构化标注。 引入可验证的多任务奖励系统,联合优化文档解析、布局分析、公式解析等八个目标。 发布 Flash 和 Pro 两个变体,Pro 版本在 olmOCR-Bench 和 ParseBench 上达到 SOTA 性能。

65
Hot 热度
75
Quality 质量
70
Impact 影响力

Analysis 深度分析

TL;DR

  • Infinity-Parser2 introduces a novel architecture combining a controllable data-synthesis engine with multi-task reinforcement learning to solve the scarcity of high-quality document parsing data.
  • The team released Infinity-Doc2-5M, a 5-million-sample bilingual corpus featuring precise annotations for bounding boxes, reading order, and various content formats like Markdown, LaTeX, and chemical structures.
  • A verifiable, multi-task reward system enables Joint Reinforcement Learning across eight distinct objectives, unifying perception, structural analysis, and reasoning into a single optimization signal.
  • Two model variants are available: Infinity-Parser2-Flash for high-throughput, low-latency inference (3.68x faster than its predecessor) and Infinity-Parser2-Pro for maximum precision.
  • Infinity-Parser2-Pro achieves state-of-the-art results on olmOCR-Bench (87.6%) and ParseBench (74.3%), outperforming competitors like DeepSeek-OCR-2 and MinerU2.5, particularly in complex domains like charts and chemical formulas.

Why It Matters

This development addresses a critical bottleneck in document AI: the lack of faithful, large-scale annotated data. By automating data synthesis and leveraging reinforcement learning for multi-task alignment, Infinity-Parser2 offers a scalable path to high-precision document understanding without relying solely on manual annotation. For practitioners, the availability of both a speed-optimized and precision-optimized variant allows for flexible deployment strategies depending on latency versus accuracy requirements in production environments.

Technical Details

  • Data Synthesis Engine: Utilizes a controllable rendering framework with an iterative refinement loop to generate synthetic training data, resulting in the open-source Infinity-Doc2-5M corpus. This dataset includes bilingual (Chinese/English) samples with detailed annotations for layout, reading order, and specific content types (Markdown, HTML, LaTeX, SMILES, structured charts).
  • Multi-Task Reinforcement Learning: Implements a verifiable reward system that supports Joint Reinforcement Learning across eight co-trained objectives: document parsing, layout analysis, table parsing, math formula parsing, chart parsing, chemical formula parsing, document VQA, and general multimodal understanding.
  • Model Variants:
    • Infinity-Parser2-Flash: Optimized for inference speed, delivering a 3.68x throughput increase compared to Infinity-Parser-7B.
    • Infinity-Parser2-Pro: Designed for precision-critical applications, focusing on accuracy across diverse document structures.
  • Performance Benchmarks: The Pro variant sets new standards with 87.6% accuracy on olmOCR-Bench and 74.3% on ParseBench, demonstrating superior generalization to specialized domains such as chemical formulas and complex charts.

Industry Insight

The shift towards synthetic data generation coupled with reinforcement learning signals a move away from expensive, manual annotation pipelines for specialized vision-language tasks. Organizations should consider adopting similar controllable synthesis frameworks to scale their document processing capabilities. Furthermore, the dual-release strategy of Flash and Pro models highlights the importance of tailoring model selection to specific operational constraints, suggesting that future enterprise deployments may require hybrid approaches or dynamic routing between speed and precision models.

TL;DR

  • Infinity-Parser2 提出了一种结合可控数据合成管道与多任务强化学习的端到端文档解析大模型方案。
  • 构建了开源的 Infinity-Doc2-5M 数据集,包含500万双语样本,涵盖多种文档类型及结构化标注。
  • 引入可验证的多任务奖励系统,联合优化文档解析、布局分析、公式解析等八个目标。
  • 发布 Flash 和 Pro 两个变体,Pro 版本在 olmOCR-Bench 和 ParseBench 上达到 SOTA 性能。

为什么值得看

该报告展示了如何通过大规模合成数据和强化学习解决文档解析领域标注数据稀缺的核心痛点。其多任务联合优化策略为构建通用型文档理解模型提供了新的技术范式,具有极高的工程参考价值。

技术解析

  • 数据合成引擎:开发了可控渲染框架配合迭代精炼循环,构建了包含边界框、标准内容形式(Markdown/HTML/LaTeX等)及阅读顺序的500万双语数据集。
  • 多任务强化学习:设计了一套可验证的多任务奖励机制,将感知、结构和推理统一在单一优化信号中,同时训练文档解析、表格解析、化学式解析等八项任务。
  • 模型变体与性能:Infinity-Parser2-Pro 在 olmOCR-Bench 达到 87.6%,ParseBench 达到 74.3%,超越了 DeepSeek-OCR-2 等竞品;Infinity-Parser2-Flash 吞吐量提升 3.68 倍,侧重低延迟推理。

行业启示

  • 数据策略转变:从依赖人工标注转向“可控合成+迭代精炼”的大规模自动化数据生产模式,是突破垂直领域数据瓶颈的关键路径。
  • 多任务联合优化:通过单一奖励信号统一多个相关子任务,能有效提升模型的泛化能力和资源利用率,避免碎片化的模型开发。
  • 场景化模型分发:针对精度敏感和延迟敏感场景提供不同变体(Pro/Flash),体现了大模型落地中平衡性能与成本的重要工程实践。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Multimodal 多模态 Research 科学研究 Dataset 数据集 Training 训练