Do language models miss presuppositions that arise from context?

Presuppositions come from two sources: fixed word meanings and conversational dynamics. Can LLMs that learn trigger patterns detect presuppositions that emerge from discourse accommodation rather than lexical items?

Note · 2026-02-21 · sourced from Natural Language Inference

Formal semantics distinguishes two routes by which presuppositions enter discourse:

Lexical specification: Certain lexical items conventionally carry presuppositions as part of their meaning. "John stopped smoking" presupposes John was smoking — this is encoded in the lexical semantics of stop. The presupposition is stable across contexts: it survives embedding, negation, and questioning in predictable ways.
Conversational derivation: Some presuppositions are not encoded in any trigger but arise from conversational dynamics — specifically, from accommodation. When a speaker asserts "The present king of France is wise," the listener accommodates the presupposition that there is a present king of France to keep the conversation coherent. This presupposition was not triggered by any specific lexical item; it emerged from the structure of the discourse.

LLMs learn statistical associations between trigger lexemes and the inference patterns they generate. This gives them systematic but incomplete coverage: they can handle lexically-specified presuppositions (at least in simple embedding contexts) but they fail at conversationally derived presuppositions because those require:

Tracking the Question Under Discussion (QUD) — what are we currently trying to resolve?
Understanding what counts as discourse-new vs. discourse-given in this conversation
Performing accommodation — the dynamic updating of shared context to make the incoming content coherent

Accommodation is the key mechanism. It is not triggered by a word; it is triggered by a mismatch between what the discourse assumes and what the speaker's utterance requires. LLMs that have learned trigger patterns will miss these accommodations because they are looking for lexical hooks that don't exist.

This is an extension of Does projection strength vary by context or by word type?: the Gradient Projection Principle shows that even lexically-triggered presuppositions vary in projection strength based on discourse context. Conversationally derived presuppositions add a second layer of context-sensitivity that goes beyond even the gradient revision of trigger-based accounts.

The implication for Why do embedding contexts confuse LLM entailment predictions? is that the failures observed there (LLMs treating triggers and non-factives identically) represent only one dimension of the problem. Even if LLMs were fixed to correctly handle lexical triggers, they would still fail at conversationally derived presuppositions — a harder problem that requires genuine discourse tracking rather than pattern recognition.

Source: Natural Language Inference

Related concepts in this collection

Does projection strength vary by context or by word type? Standard accounts treat presupposition projection as categorical, but do English expressions actually project uniformly? This question explores whether context and discourse role determine how strongly content survives embedding.
gradient projection applies to lexical triggers; conversational derivation adds a further context-sensitivity layer
Why do embedding contexts confuse LLM entailment predictions? Can language models distinguish between contexts that preserve versus cancel entailments? The study explores whether LLMs systematically fail to apply the semantic rules governing presupposition triggers and non-factive verbs.
that paper shows LLM failure on lexical triggers; this note explains why the problem extends further
Why do language models fail at communicative optimization? LLMs excel at learning surface statistical patterns from text but struggle with deeper principles of how language achieves efficient communication. What distinguishes these two types of linguistic knowledge?
trigger learning is statistical; accommodation requires communicative reasoning
Why do language models skip the calibration step? Current LLMs assume shared understanding rather than building it through dialogue. This explores why that design choice persists and what breaks when it fails.
accommodation is a dynamic grounding mechanism; LLMs that presume common ground cannot accommodate

Concept map

14 direct connections · 147 in 2-hop network ·dense cluster

Do language models miss presuppositions that ari… Does projection strength vary by context or by wor… Why do embedding contexts confuse LLM entailment p… Why do language models fail at communicative optim… Why do language models skip the calibration step?

Click a node to walk · click center to open · click Open full network for a force-directed map

your link semantically near linked from elsewhere

Original note title

presuppositions have a dual origin — lexical specification and conversational derivation — and llms that learn trigger patterns miss conversationally derived presuppositions