Hyperbots Presents SAVIOR at WACV 2026

Hyperbots Inc., a leading Agentic AI platform for Finance & Accounting, today announced the publication of its new research paper, “SAVIOR: Sample-efficient Adaptation of Vision-Language Models for OCR Representation,” showcased at the WACV 2026 Workshop, part of one of the world’s leading computer vision conferences. The research introduces SAVIOR, a targeted data curation framework that improves OCR accuracy for complex financial documents using a surprisingly small but carefully selected training dataset.

OCR continues to be a major bottleneck in finance operations. While modern Vision-Language Models perform well on standard benchmarks, they often struggle with the types of documents commonly seen in enterprise workflows. These include invoices with rotated labels, logo-heavy vendor headers, fine-print compliance text, degraded scans, and dense multi-column layouts. Although these edge cases are rarely represented in public datasets, they account for a large share of real-world OCR failures across invoices, receipts, statements, and compliance documents.

SAVIOR approaches the problem differently. Instead of relying on massive amounts of generic training data, the framework focuses on curating documents that represent the exact visual structures where pretrained models tend to fail. The resulting training dataset, SAVIOR-Train, contains just 2,234 documents selected specifically for these challenging patterns.

Using SAVIOR-Train, Hyperbots fine-tuned Qwen2.5-VL-Instruct and achieved a word-level recall of 0.93. When the OCR output was used for downstream document question answering tasks, the model achieved an F1 score of 0.89, outperforming systems such as GPT-4o, Mistral-OCR, PaddleOCR-VL, and DeepSeek-OCR despite using a fraction of the training data.

The research also introduces SAVIOR-Bench, a 509-document benchmark built from real-world finance and accounting workflows. In addition, the paper proposes PaIRS (Pairwise Relational Similarity), a structure-aware evaluation metric designed to measure whether OCR systems preserve spatial relationships between words and fields, something traditional text-based metrics often miss.

“Most OCR benchmarks don’t reflect what actually breaks in enterprise finance workflows,” said Niyati Chhaya, Co-Founder & VP AI at Hyperbots Inc. “SAVIOR shows that carefully curated data targeting real production failure modes can outperform much larger systems trained on generic datasets. This is how we approach AI at Hyperbots: focused on production accuracy, operational reliability, and practical deployment.”

The research also demonstrated that SAVIOR-trained models maintain strong performance even under 4-bit quantization, making them suitable for memory-constrained and latency-sensitive enterprise environments. The quantized 3B model achieved throughput exceeding 1,085 tokens per second on an A100 GPU, enabling scalable deployment in high-volume document pipelines.

The ideas behind SAVIOR directly support Hyperbots’ broader AI architecture for finance and accounting automation. Hyperbots’ AI Co-Pilots help enterprises automate invoice processing, accruals, vendor management, procurement, payments, tax verification, cash reconciliation, collections, forecasting, and reporting workflows using finance-specific AI systems designed for enterprise accuracy and control.

By publishing SAVIOR, Hyperbots continues its investment in production-focused AI research for enterprise finance. The company’s AI team holds more than 50 patents across artificial intelligence, document understanding, and enterprise software.

Full Research Paper: https://openaccess.thecvf.com/content/WACV2026W/VisionDocs/papers/Bhat_SAVIOR_Sample-efficient_Adaptation_of_Vision-Language_Models_for_OCR_Representation_WACVW_2026_paper.pdf

About Hyperbots Inc.

Hyperbots Inc. transforms Finance & Accounting operations using proprietary AI agents that autonomously read, reconcile, validate, and post financial data across complex workflows. Its specialized AI Co-Pilots operate on a finance-specific foundational architecture delivering intelligence across language, vision, reasoning, interpretation, recommendation, prediction, redaction, and exception handling. Hyperbots integrates with major ERPs and maintains ISO 27001, SOC 1 Type 2, and SOC 2 Type 2 certifications.

Contact: support@hyperbots.com
Website: https://www.hyperbots.com
Request a Demo: https://www.hyperbots.com/request-demo

About the WACV 2026 Workshop

The IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) is one of the leading peer-reviewed conferences in computer vision research. Its workshops focus on applied and emerging research with an emphasis on real-world deployment and scalability.