Can LLMs compute how presuppositions project through embedded clauses?

This explores whether LLMs can track presupposition projection — how a presupposition triggered deep inside a sentence survives, or gets cancelled, depending on what embeds it. The corpus is fairly blunt: by default, no. The most direct evidence is that embedding contexts act as systematic 'blinds.' LLMs treat presupposition triggers and non-factive verbs as surface cues rather than computing how those verbs flip the semantic effect on what's entailed — and this failure is stable across prompts and models, not a fluke of phrasing Why do embedding contexts confuse LLM entailment predictions?. The deeper reason shows up when you decompose where presuppositions even come from: some are lexically specified by trigger words (which pattern-matching can catch), but others arise through accommodation — quietly updating context to resolve a discourse mismatch — and tracking those requires modeling the questions under discussion, not matching tokens Do language models miss presuppositions that arise from context?.

Projection is precisely the hard case, because it's structural: the answer depends on syntactic depth and the operators stacked above the trigger. And LLMs degrade predictably as that structure deepens — top models consistently misidentify embedded clauses, verb phrases, and complex nominals, which is the raw material projection runs on Why do large language models fail at complex linguistic tasks?. The failure is compounded by what models do instead of inference: entailment judgments lean on whether the hypothesis looks attested in training data rather than whether the premise actually supports it Do LLMs predict entailment based on what they memorized?, and more broadly, when you decouple semantic content from logical form, performance collapses even with correct rules supplied in context Do large language models reason symbolically or semantically?. Projection demands exactly the symbolic bookkeeping this suggests they lack.

There's a striking adjacent finding that sharpens the diagnosis: models will accept false presuppositions even when they demonstrably know the underlying fact is wrong — rejection rates fall far below knowledge — so the gap isn't missing knowledge but a missing step that applies it Why do language models accept false assumptions they know are wrong?. That's the same shape as 'Potemkin understanding,' where a model can correctly explain a concept, fail to apply it, and even recognize the failure — a sign of disconnected explanation and execution pathways rather than a simple gap Can LLMs understand concepts they cannot apply?.

The hopeful twist — and the thing you might not expect — is that the capability seems to exist but stays dormant. When pushed into explicit step-by-step reasoning, models like o1 can build genuine syntactic trees and metalinguistic generalizations, going well beyond surface behavior Can language models actually analyze language structure?. And forcing structured reasoning steps — making a model check warrants and implicit premises instead of skipping them — catches failures that ordinary prompting waves through Can structured argument prompts make LLM reasoning more rigorous?. So the honest answer is layered: LLMs can't compute projection reliably in their default fast mode, where it reduces to pattern-matching triggers, but the structural analysis may be recoverable when you explicitly scaffold the inference. Whether that scaffolded competence is robust or just another well-explained-but-misapplied surface remains the open question.