Representational Dynamics of Pragmatic Processing

Instruction-tuned language models often extract the meaning of a sentence based on the overall context. For instance, if you invite somebody to go out and that person replies with "I start work early tomorrow", the model may behave in a pragmatic way and infer that the person is refusing the invitation. However, it remains unclear whether this ability is linked to representational signatures characteristic of pragmatic processing. Alternatively, both pragmatic and non-pragmatic behavior may share similar internal representations, and the difference in output behavior would merely emerge from context scaffolding. To shed light on LLM-based pragmatic inference, we evaluate a panel of 44 instruction-tuned models (1--34B) on 648 Spanish two-turn exchanges in which a polar question (element 1) leads to an indirect answer (element 2) that must be matched either to its implicature ($\mathrm{I}$, element 3) or to a literal paraphrase ($\mathrm{PU}$, element 4), under 0 to 4 in-context examples. Behaviorally, model size is a weak predictor of pragmatic responding at 0 shots but a strong one at 4 shots: scale mainly buys the capacity to exploit demonstrations. We then ask whether that uptake is accompanied by pragmatic-specific reorganization of the representations of the four elements composing each exchange, using within-element cross-layer CKA and within-span representational curvature. Aggregate change in cross-layer similarity does not track pragmatic uptake; change in curvature does, and this effect concentrates on $\mathrm{I}$, the element carrying the inference. Our results suggest that pragmatic gains correspond to a specific, local geometric change rather than broad representational changes.