SAGE framework blends LLMs and cognitive models for pragmatic reasoning

Pragmatic language use—how we choose words and interpret meaning—relies on reasoning about alternatives: what else a speaker could have said, or what else a listener might understand. Existing computational models often manually specify these alternatives, limiting their flexibility. A new paper on arXiv introduces SAGE (ScAffolded Generative models for Explanation), a framework that fuses the transparency of cognitive models with the generative power of large language models (LLMs). SAGE breaks pragmatic reasoning into three modular components: proposers that use LMs to generate open-ended candidate alternatives; evaluators that assess those alternatives on dimensions like semantics, complexity, or typicality; and selectors that apply rule-based, cognitively motivated steps to pick a final output. This neuro-symbolic design aims to retain explainability while scaling to real-world linguistic variation.

The authors tested SAGE across three case studies: referential expression generation (e.g., picking a unique name for an object), manner (M-)implicatures (e.g., why saying “He sat down” vs. “He plopped down” conveys different attitudes), and Gricean conversational implicatures (e.g., how listeners infer meaning beyond literal statements). SAGE models achieved high accuracy and often outperformed baseline methods in quantitative fits to human data. However, component-level analyses revealed a key asymmetry: LM-powered proposers reliably generated diverse, pragmatically relevant alternatives, while LM-based evaluators proved stronger at providing intuitive judgments than at measuring formal or theoretical properties. The researchers see this as both a promise (for flexible generation) and a limitation (for faithful cognitive modeling), highlighting the ongoing challenge of building explanatory accounts of human pragmatics.