Tag
This paper investigates whether large language models exhibit pragmatic cooperativity in multi-party collaborative tasks under conditions of epistemic asymmetry, formalizing Grice's cooperative principle and evaluating LLMs as both speakers and listeners. Results show that while LLMs display some pragmatic capabilities, they struggle with incomplete information and fail to recognize certain violations of Gricean maxims.
PragReST is a self-supervised framework that improves LLM pragmatic reasoning by generating counterfactual reasoning traces and training models via supervised fine-tuning and reinforcement learning, achieving significant gains on pragmatic benchmarks without human-labeled data.