Large tabular models explained: Understanding a new form of gen AI

A large tabular model is a type of gen-AI system that can analyze large-scale tables and other kinds of organized datasets, such as billions of payments transactions stored neatly in fixed rows and columns. It can look for patterns and make predictions using these datasets. In this way, it can help spot trends and behaviors that would be hard for people, or other types of software, to uncover. 

Large language models, which power OpenAI’s ChatGPT and Google’s Gemini, are trained on text, images and videos, also known as unstructured data. They’re especially good at creating similar types of content, such as short stories or images, and can also be used to develop software code.

An LTM is trained on structured data, like the rows and columns of transaction records. LTMs are good at finding patterns, predicting behaviors and potentially improving systems behind the scenes.

Both types of generative AI models are trained using huge volumes of data — whether that’s written text for LLMs or organized datasets for LTMs. Both types of models can then use this training to analyze and predict.

For an LLM, that can mean predicting what the next word in a sentence may be after it analyzed millions and millions of sentences. That’s how it’s able to write a new story or song lyrics.

For an LTM, that can mean predicting someone's next payment transaction after it analyzed billions of prior transactions. That way, it can inform fraud prevention models on what is likely to be a legitimate transaction and what might be fraud.

While LLMs aren’t as strong with structured data, LTMs tend to excel at processing and analyzing this kind of information. Because LTMs and LLMs learn from different types of data, one doesn’t replace the other. Rather, they do different things and can complement one another. 

LTMs can be used to analyze different datasets and structured data, allowing users to find new connections and gain new insights across large and disparate tables with many rows of data.

In financial services and payments, an LTM could help spot suspicious activity earlier and more accurately by learning patterns from large amounts of transaction data. It could recognize typical spending patterns and quickly identify when something suddenly doesn’t fit that pattern. Because an LTM learns what normal behavior looks like across many situations, it can help fraud systems also avoid flagging legitimate purchases as fraud, reducing unnecessary card declines while still looking for real risk.

Mastercard started developing its own LTMs, trained on billions of anonymized transactions, to investigate exactly these kinds of scenarios and uses. We believe our LTMs could be used for cybersecurity, loyalty and rewards programs, personalization models, portfolio optimization and data analytics tools. Additionally, because an LTM is more flexible than existing traditional machine-learning AI systems, we may be able to use a handful of LTMs instead of building and maintaining thousands of more traditional machine learning models today.

In addition to payments and financial services, LTMs could be used for any industry that manages large amounts of structured data — from pharmaceuticals to energy to transportation and logistics — so organizations can identify new insights in their data.

Based on publicly available information as of the publication date of this article, it appears that there are few — if any — LTMs in production today. However, many companies have announced positive results so far from their research and development of their own LTMs, and many appear to be working toward production. News is coming out every day about generative AI capabilities, so rapid changes and new discoveries involving LTMs should be expected.

For now it appears that, unlike LLMs, LTMs are more likely to operate behind the scenes, strengthening fraud prevention systems, analyzing large datasets and strengthening data science capabilities within organizations.

We don’t yet know how LTMs could interact with AI agents. Today, LLMs are used as the foundation to power many kinds of AI agents, like ChatGPT, but the same kind of use cases aren’t as obvious for LTMs since they are being used more to improve and streamline organizations’ operations and systems. However, we can imagine a future in which an agent could work with an LTM to solve real-world problems in a similar fashion to how that’s done today with LLMs.

For example, an agent could someday use the analysis from an LTM to power personalized recommendations for someone shopping through a gen-AI chatbot. In this way, LTMs could eventually be used to help power agentic commerce, a new and developing form of retail using these agents to shop, discover and buy.