What is a Large Causal Model? Reconciling three competing definitions

MIT Sloan Management Review recently concluded that machine learning's insights into what a business choice will actually cause are flawed, at best, because ML relies on correlations, not causes, to make predictions.

Predicting what happens next is a different task from estimating what a price change, a media shift, or a supply reshuffle will actually cause, and causal modeling is the analytical lever built for that second task.

That distinction has a naming problem. Since December 2025, three separate sources, two peer-reviewed or preprint papers and one public synthesis, have used the exact phrase "Large Causal Model" for three different constructs, and DATA POEM uses the same phrase for a fourth. A buyer comparing vendor claims cannot tell which one they are being sold without knowing all four.

A Large Causal Model is a foundation model trained to represent cause-and-effect relationships in a domain's data. That is a different training objective from the word-correlation patterns a large language model learns from text. It is used to estimate what a specific input change would do to a specific outcome, a question prediction alone cannot answer.

It differs from an LLM in what it optimizes for: an LLM predicts the next plausible word; a Large Causal Model estimates an effect.

Key takeaways

• A Large Causal Model is a foundation model trained on cause-and-effect relationships. An LLM is trained on the word-correlation patterns it learns from text instead.
• Since December 2025, three separate sources, two peer-reviewed or preprint papers and one public synthesis, have used "Large Causal Model" for three distinct constructs, alongside DATA POEM's own use of the term.
• DATA POEM's Large Causal Model, FOUNT, is closest to the purpose-built causal foundation model framing below, not the LLM-extraction or LLM-augmentation framings.
• The construct underneath a vendor's "Large Causal Model" claim determines what data grounds it, and what kind of question it can actually answer.

Why the definition is suddenly contested

Most people assume "Large Causal Model" is a single, settled term, coined once and used consistently since. It is not. The reason is structural.

A NAACL survey had already established causal inference with LLMs as an active subfield before the term existed. When multiple groups needed a name for what they were building on top of that subfield, more than one reached for the same two words.

Mahadevan's DEMOCRITUS paper used it for a causal-knowledge structure extracted from an existing LLM.

Kougioulis and colleagues' temporal causal discovery paper used it for a purpose-built causal foundation model.

A synthesis on Emergent Mind, drawing on public commentary rather than original research, frames it as an LLM augmented with causal-reasoning modules.

All three shape how the market talks about the term, but none describes what DATA POEM built.

How POEM365’s Large Causal Model works

FOUNT is our Large Causal Architecture, and POEM365 is the Large Causal Model built on it. FOUNT is grounded in causal inference, the framework Judea Pearl developed to formalize cause and effect, and pretrained across cross-functional business data, marketing, finance, supply, and operations combined. Each client deployment fine-tunes that pretrained base into an Enterprise Growth Model, adapted to that company's own data and structure.

The mechanism matters because of what it replaces. Traditional marketing mix modeling is built on ordinary least squares regression, which estimates the marginal effect of one variable at a time. Add a second business function, and the assumptions break down, what practitioners call the Gordian Knot of Multicollinearity.

A Large Causal Model solves this by representing the causal graph, the actual structure of what drives what, so a change in one function's inputs is modeled as a cause with a traceable effect elsewhere rather than one more correlated variable competing for the same signal.

Three ways "Large Causal Model" gets used, and where DATA POEM's differs

Strip away the shared name, and three distinct constructs remain. Here is what each means, and how FOUNT compares to them.

A Large Causal Model as a knowledge graph extracted from an LLM

Mahadevan's December 2025 preprint introduces DEMOCRITUS, a system that extracts a causal-knowledge structure from an already-trained LLM. The construct mines causal relationships implicit in what the LLM absorbed during training, instead of training a new model directly on a causal objective.

POEM365 is pretrained directly on business data, with causal inference as the objective from the outset, rather than extracted from an existing LLM's latent knowledge. The distinction matters: an extracted graph is only as good as the causal patterns the underlying LLM happened to pick up from whatever text it saw.

A Large Causal Model as a purpose-built foundation model

Kougioulis and colleagues propose Large Causal Models as their own class of foundation model, purpose-built for temporal causal discovery in time-series data, trained directly for causal structure rather than derived from a language model.

This is the closest of the three to DATA POEM's own usage. FOUNT is likewise purpose-built for causal representation from the outset. The difference is domain: Kougioulis and colleagues target general time-series discovery; FOUNT is trained on enterprise growth data and fine-tuned per client.

A Large Causal Model as an LLM with causal-reasoning modules added

Emergent Mind's synthesis frames the Large Causal Model as a pretrained LLM with causal-reasoning modules added on top. The language model still does the underlying representation work; a causal-reasoning layer is bolted on to answer counterfactual questions the base model could not otherwise handle.

FOUNT differs here too. Causal inference is its base objective from the start; no language-model layer sits underneath its causal estimates, and no causal-reasoning module gets added to a general-purpose LLM afterward.

Why the distinction matters when evaluating vendor claims

For a CFO or CMO evaluating a vendor's "Large Causal Model" claim, the construct underneath it changes what you can trust the model to do.

A model derived by extracting causal knowledge from an already-trained LLM inherits whatever causal patterns that LLM picked up from its training text. Ask what grounds the causal claim: a text corpus, or your transactions and spend.

A model built as an LLM with causal-reasoning modules bolted on still has a language model doing the underlying representation work, with a reasoning layer generating "what if" answers on top. Ask whether that answer is empirically estimated, or produced the way an LLM produces any other plausible-sounding text.

A purpose-built causal foundation model, trained directly on your operational data with causal inference as its base objective, is the only one of the three built to answer a counterfactual with the rigor a controlled experiment would. That is the test worth applying before you sign anything.

Common misconceptions about Large Causal Models

A few assumptions get in the way here

• "Large Causal Model" means the same thing everywhere. It does not. DATA POEM's own use is closest to the purpose-built foundation model framing above.
• Any AI that discusses causality is a Large Causal Model. Mentioning causal language in an output is not the same as being trained with a causal objective. An LLM prompted to "explain the cause" is still predicting the next plausible word.
• A Large Causal Model is a variant of a large language model. A Large Causal Model is built on the foundation of causal AI, and its input is time series data, not the text or images an LLM trains on. In POEM365's case there is no LLM layer underneath its causal estimates. 
• Peer-reviewed use of the term means it is standardized. Three groups using the same two words for three different systems within a single year is the opposite of standardization.

FAQs

Is "Large Causal Model" a peer-reviewed or standardized term?

"Large Causal Model" is not yet a standardized, peer-reviewed term with one settled definition. Since December 2025, at least three peer-reviewed or preprint sources, Mahadevan's DEMOCRITUS paper, Kougioulis and colleagues' temporal causal discovery paper, and a NAACL 2025 survey on the broader field, have used "Large Causal Model" or closely related language for distinct constructs, none identical to DATA POEM's own use.

Is DATA POEM's Large Causal Model the same construct described in recent AI research papers?

DATA POEM's Large Causal Model, POEM365, is not exactly the same construct described in recent AI research papers. It is closest to the purpose-built causal foundation model framing used by Kougioulis and colleagues: built directly on causal structure rather than extracted from, or bolted onto, an existing LLM. POEM365 is the model, built on FOUNT, our Large Causal Architecture. It differs in domain, since it is trained specifically on enterprise growth data rather than general time series.

What is causal AI?

Causal AI is AI built to estimate cause-and-effect relationships, what a specific input would change, not to predict the most statistically likely outcome. It answers counterfactual questions a purely predictive model cannot. Our full breakdown of causal AI covers the underlying mechanics in more depth.

How does causal AI work?

Causal AI works by modeling the causal graph, the actual structural relationships between variables, instead of fitting a correlation between them. It uses that structure to estimate counterfactuals: what would have happened to an outcome if one input had changed, with everything else held constant.

How does a Large Causal Model differ from a large language model?

A Large Causal Model estimates the effect of changing a specific input on a specific outcome; a large language model predicts the statistically likely next word in a sequence instead. For the full tool-choice comparison, see Pick the Right AI for the Job: LLM vs. LCM.

Which Large Causal Model are you actually being sold?

The three constructs above are evidence of how fast "Large Causal Model" became a real technical category, faster than the industry could agree on what belongs inside it.

I expect the ambiguity to widen before it narrows. As more vendors borrow the term for anything from a causal-reasoning add-on to a purpose-built foundation model, the two words will do less work on their own, and asking what is underneath them will matter more.

At DATA POEM, FOUNT is our purpose-built Large Causal Architecture, built from the ground up on causal inference with no LLM layer added on, and POEM365 is the Large Causal Model that runs on it. If you are trying to work out what is actually running under a vendor's "Large Causal Model" claim, see how POEM365 puts a Large Causal Model to work.