Can LLMs identify the hidden assumptions that make arguments work?

Toulmin's argument model distinguishes claim, data, and warrant. The claim is what is being argued. The data is the evidence. The warrant is the often-unstated principle connecting data to claim — the implicit assumption that makes the inference valid.

In natural language, warrants are almost never stated. When someone argues "This policy failed in Europe, so it will fail here," the unstated warrant is something like "contexts similar to Europe will produce similar outcomes." Evaluating the argument requires identifying this warrant and assessing its validity in context.

The Argument Reasoning Comprehension task tests this capability directly. LLMs perform well on identifying the explicit claim-data structure — recognizing what is being argued and what evidence is offered. They fail significantly at supplying or evaluating the implicit warrant. The gap between structural recognition and warrant identification is large.

This is a different failure than Why does ChatGPT fail at implicit discourse relations?. That finding concerns discourse relations (because, therefore, although). This finding concerns argumentative inference — the background knowledge required to evaluate whether data actually supports a claim. Both are implicit-structure failures, but at different levels.

The failure is not simply about world knowledge being absent. Do language models actually use their encoded knowledge? suggests relevant knowledge may be encoded but not accessed when needed. Warrant identification requires activating world knowledge in response to argumentative context — a different retrieval trigger than direct factual recall.

Practically: LLMs can generate the surface form of argumentation (claim, evidence, conclusion) without the inferential work that makes the argumentation valid. They can look like they are reasoning about arguments without engaging with the warrants that determine whether the arguments hold.