- Matrix streamlines document processing by cutting manual labor and operational costs, using AI agents in the enterprise.
Today, Beyond Work, an enterprise AI company, announced the record-setting results of Matrix, a novel memory-augmented AI framework for automating business document processing. Developed in collaboration with researchers from Penn State University, Oregon State University, and Kuehne+Nagel, one of the world’s largest logistics providers, Matrix addresses the complex, time-intensive task of extracting transport references from Universal Business Language (UBL) invoices.
Comparing the success rates of four methods (CoT, Two-agent, Reflexion, Matrix) across GPT-4o-mini and GPT-4o, with Matrix achieving the highest performance.
By harnessing an iterative, memory-centric learning strategy, Matrix achieves a 30.3% improvement over chain-of-thought prompting, outperforms a standard Large Language Model agent by 35.2%, and surpasses Reflexion by 27.28%-establishing its state-of-the-art capabilities in AI reflection.
“Matrix redefines what’s possible for enterprise automation by dramatically improving accuracy while reducing operational costs,” said Malte Højmark Bertelsen, co-author and cofounder of Beyond Work.
Matrix’s success is the result of an international team of experts, including Jiale Liu, Yifan Zeng, Malte Højmark-Bertelsen, Marie Normann Gadeberg, Huazheng Wang, and Qingyun Wu, an Assistant Professor at Penn State University recognized for her contributions to Automated Machine Learning (AutoML) and Large Language Models (LLMs). Her track record includes high-impact open-source projects, such as AutoGen, that enable complex multi-agent collaborations – foundational principles driving Matrix’s memory-augmented approach.
Key Highlights
- Real-World Validation: Data from Kuehne+Nagel demonstrates Matrix’s impact on global logistics operations.
- Iterative Learning: Self-reflection accelerates domain adaptation for specialized documents.
- Operational Efficiency: Fewer API calls and reduced cost profile elevate enterprise scalability.
- Enhanced Robustness: The system effectively handles larger, more complex documents beyond typical AI baseline models.
An anonymized subset of the dataset is available to catalyze further research in enterprise AI by contacting Beyond Work.
Research Reference
Paper: https://arxiv.org/abs/2412.15274
Open-source data: https://github.com/bwllaming/matrix-paper
About Beyond Work
Co-founded by industry veterans from Uber, Tradeshift, and other unicorn alumni, Beyond Work is an enterprise AI platform that eliminates tedious tasks and drives tangible business outcomes in finance, procurement, and supply chain. Used by Fortune 500 customers in energy, logistics, and life sciences, its state-of-the-art platform leverages agentic networks in business to empower teams to focus on real innovation instead of busy work.
Contact Information
Malte Højmark-Bertelsen
Cofounder, Head of Applied AI and Research
malte@beyondwork.ai
SOURCE: Beyond Work