Pragmatics
How context and intention determine what is communicated beyond the literal words — the level where most prompt frustration actually lives, and the one models handle least evenly.
When not to use it
- As an explanation for failures that are representational, such as character counting, where the information was destroyed before the model saw it.
- To excuse an underspecified prompt. If the instruction was ambiguous to a person too, that is not a pragmatics failure.
Reach for something else instead
- Explicit instruction stating the speech act, quantity and constraints directly.
- Structured output formats that remove the inference entirely.
- Few-shot examples demonstrating the intended reading.
Sources & further reading
- Grice (1975), Logic and Conversation — the maxims and the cooperative principle.
- Hu et al. (2023), A fine-grained comparison of pragmatic language understanding in humans and language models — the conventional versus novel split.
- Ruis et al. (2023), The Goldilocks of Pragmatic Understanding — implicature performance across model scales and tuning.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Assuming politeness is neutral. Indirectness is how politeness works in English, and it is exactly what fails.
- Reading strong benchmark results as competence, since most measure recognition under instruction rather than pragmatic behaviour in use.
- Expecting a better model to fix it, when the evidence for monotonic improvement is weak.
At a glance
Where this sits
A destination. 1 concept lead here, and nothing in the corpus depends on it.
Computed from the prerequisite graph, not assigned. How this works