Existential presupposition is a foundational component of meaning: it reflects implicit assumptions of existence that underlie interpretation, even when not explicitly stated. Sentences such as "Neo bends the spoon" presuppose that the entities referred to exist, independent of the truth-value of the sentence itself. Because this type of meaning is implied rather than explicitly asserted, it provides a diagnostic test of whether large language models (LLMs) display sensitivity to more abstract and less surface-driven layers of meaning. We adapt a natural language inference (NLI)–based probing setup, using a fine-tuned version of DeBERTa-v3-large as a baseline model and compare its behaviour to that of LLaMA-3.1-8B-Instruct and Gemma-3-12B-it under zero- and few-shot prompting, as well as to their fine-tuned base-variants. We find that while all models show sensitivity to existential presupposition across syntactic embeddings, determiner types and contextual cues, their behaviour differs markedly in strength and systematicity, with NLI-fine-tuned autoregressive models exhibiting the most coherent and stable projection patterns. They showed graded and theoretically aligned projection patterns, whereas instruction-tuned models remain largely prone to surface heuristics and prompt susceptibility. These results suggest that pre-trained LLMs exhibit sensitivity to existential presupposition but this behaviour surfaces only systematically when the models have learned the intricacies of the NLI task.

Top