FireRedTeam Releases FireRed-OCR-2B Utilizing GRPO to Solve Structural Hallucinations in Tables and LaTeX for Software Developers

Document digitization has long been a multi-stage problem: first detect the layout, then extract the text, and finally try to reconstruct the structure. For Large Vision-Language Models (LVLMs), this often leads to ‘structural hallucinations’—disordered rows, invented formulas, or unclosed syntax.

The FireRedTeam has released FireRed-OCR-2B, a flagship model designed to treat document parsing as a structural engineering task rather than ‘impressionist’ text generation. Built on the Qwen3-VL-2B-Instruct architecture, this model establishes a new State-of-the-Art (SOTA) for end-to-end solutions, achieving an overall score of 92.94% on the OmniDocBench v1.5 benchmark.

Shifting the Paradigm: Structural Engineering vs. Text Generation

Devs often find that even the most powerful general VLMs struggle with the dense spatial logic of a technical PDF. When a model ‘sees’ a complex table or a multi-line LaTeX equation, it frequently fails to maintain the hierarchical relationship between elements.

FireRed-OCR-2B addresses this through a specialized Progressive Training Pipeline consisting of three distinct stages:

1. Multi-task Pre-alignment: This stage establishes spatial grounding by training the model on detection, region recognition, and layout-to-markdown tasks.
2. Specialized SFT (Supervised Fine-Tuning): The model is fine-tuned on a high-quality, standardized Markdown dataset to ensure logical consistency and hierarchical expression.
3. Format-Constrained GRPO: The final stage uses reinforcement learning to enforce syntactic validity.

The Core Innovation: Format-Constrained GRPO

The most significant technical differentiator for FireRed-OCR is its use of Format-Constrained Group Relative Policy Optimization (GRPO). While traditional fine-tuning focuses on character accuracy, GRPO introduces a reinforcement learning loop that rewards the model for specific structural traits:

• Formula Syntax: Ensuring LaTeX equations are mathematically valid.
• Table Integrity: Maintaining consistent row/column counts and proper HTML/Markdown tagging.
• Hierarchical Closure: Verifying that all opened structural tags (like lists or headers) are correctly closed.
• Text Accuracy: Reducing character-level errors in dense text blocks.

By eliminating the need for a separate ‘critic’ model—a key benefit of the GRPO algorithm—FireRedTeam has optimized the training process to focus specifically on the high-friction areas of document parsing.

Solving the Long-Tail Layout Problem

The ‘long-tail’ of document layouts (e.g., non-standard legal forms, academic papers with overlapping figures, or handwritten annotations) is where most OCR pipelines break. FireRed-OCR utilizes a ‘Geometry + Semantics’ Data Factory.

This novel approach uses geometric feature clustering and multi-dimensional tagging to synthesize balanced datasets. By combining geometric awareness with semantic understanding, the model maintains ‘In-the-Wild Robustness,’ outperforming traditional pipeline systems like PaddleOCR on complex, non-standard layouts (benchmarked on the FireRedBench dataset).

Performance Benchmarks

In head-to-head comparisons on OmniDocBench v1.5, FireRed-OCR-2B (92.94%) significantly outperforms other end-to-end models, including:

• DeepSeek-OCR 2: 91.09%
• Gemini-3.0 Pro: 90.33%
• Qwen3-VL-235B: 89.15%

While some ‘pipeline’ solutions (which use separate models for detection and recognition) achieve slightly higher scores, FireRed-OCR-2B represents the leading performance for a single-model, end-to-end approach. This is particularly relevant for devs looking to reduce system complexity and inference latency in production RAG (Retrieval-Augmented Generation) environments.

Key Takeaways

I have summarized the technical significance and performance metrics of the FireRed-OCR-2B release into five key takeaways for AI engineers and data scientists.

5 Key Takeaways: FireRed-OCR-2B

• New End-to-End SOTA Performance: FireRed-OCR-2B has achieved a state-of-the-art (SOTA) score of 92.94% on the OmniDocBench v1.5 benchmark. This makes it the leading single-model solution for document parsing, outperforming significantly larger models like Qwen2-VL-72B and Gemini-1.5-Pro in structural accuracy.
• Architectural Foundation: Built on the Qwen2-VL-2B-Instruct (or the updated 2026 iteration) base, the model utilizes a Vision-Language-Model (VLM) approach. It replaces traditional multi-stage pipelines (separate detection, cropping, and OCR steps) with a unified, end-to-end transformer architecture that outputs structured Markdown directly.
• Structural Integrity via GRPO: A major technical differentiator is the use of Format-Constrained GRPO (Group Relative Policy Optimization). This reinforcement learning technique rewards the model for maintaining syntactic validity—specifically ensuring that LaTeX formulas, table tags, and Markdown hierarchies are logically closed and mathematically consistent.
• ‘Geometry + Semantics’ Data Factory: To solve the problem of complex ‘in-the-wild’ layouts, the FireRedTeam developed a specialized data engine. This ‘factory’ synthesizes datasets by balancing geometric layout features with semantic content, enabling the model to handle overlapping figures, multi-column academic papers, and non-standard forms more reliably than previous iterations.

Check out the Model Weight and Repo. Also, feel free to follow us on Twitter and don’t forget to join our 120k+ ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.