pragmatics

This paper proposes demographic-conditioned fusion embeddings to model perspectivist social meaning in language, showing consistent improvements over text-only baselines by integrating annotator demographics into NLP systems.

Introduces KARMA, a framework that trains a reward model on Reddit conversations to improve LLMs' context-sensitive conversational behavior via reinforcement learning, finding that the best reward model for predicting karma does not yield the best downstream alignment.

Introduces DRInQ, a benchmark for evaluating conversational implicature in question utterances, revealing that LLMs often fail to recover intended implications at inference time despite being able to generate plausible pragmatic scenarios.