Flout at Your Own Risk: LLMs Struggle with Pragmatic Cooperativity Under Epistemic Asymmetry

2026-07-13 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, quick

Summary

A novel investigation into Large Language Models' (LLMs) pragmatic reasoning capabilities in multi-party collaborative tasks under partial information conditions reveals significant challenges. Published on 2026-07-13, this research formalizes collaborative epistemic asymmetry, explicitly connecting objective task success to Grice's cooperative principle. The study empirically assesses various LLMs' abilities to act cooperatively as both speakers and listeners, employing both prompting and post-training strategies. While LLMs demonstrate certain pragmatic capabilities in collaborative settings, which can be elicited through these methods, they still struggle with pragmatic communication when information is incomplete. Specific failure modes correlate with unrecognized floutings of Grice's maxims, indicating a gap in their ability to handle nuanced cooperative communication.

Key takeaway

For NLP Engineers designing collaborative LLM agents, recognize that your models may struggle with pragmatic communication under partial information. You should explicitly account for epistemic asymmetry in multi-party systems. LLMs can fail to recognize floutings of cooperative principles. Consider incorporating specific prompting or post-training strategies to improve their cooperative abilities as speakers and listeners, mitigating miscommunication risks.

Key insights

LLMs struggle with cooperative communication under incomplete information, often failing to recognize floutings of Grice's maxims.

Principles

• Collaborative success links to Grice's cooperative principle.
• Epistemic asymmetry impacts LLM pragmatic communication.
• Prompting and post-training can elicit pragmatic capabilities.

Method

The study formalizes collaborative epistemic asymmetry and empirically assesses LLMs as speakers and listeners using prompting and post-training strategies in multi-party tasks.

In practice

• Use prompting to enhance LLM pragmatic skills.
• Consider post-training for cooperative communication.
• Design tasks to mitigate epistemic asymmetry.

Topics

• Large Language Models
• Pragmatic Reasoning
• Cooperative Communication
• Epistemic Asymmetry
• Multi-agent Systems
• Grice's Maxims

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Too Polite to Disagree: Understanding Sycophancy Propagation in Multi-Agent Systems — Providing LLM agents with peer sycophancy rankings significantly reduces sycophancy and improves discussion accuracy by 10.5%.
• From Propositional to Perceptual Asymmetry: Extending FPO to Asymmetric Partial Information Dialogue — Understanding dialogue friction in asymmetric information requires evaluating FPO from each participant's unique perspective, as an informed single viewpoint outperforms omniscient access.
• Unveiling the Limits of Large Language Models in Inferring Pragmatic Meaning from Non-Verbal Responses — LLMs significantly struggle to infer pragmatic meaning from non-verbal communication, unlike their performance with verbal cues.
• The Hidden Reason LLMs Fail in Conversations: CCOPD — CCOPD re-anchors LLM responses in multi-turn conversations to a canonical context, mitigating "self-anchored drift" from piecemeal information.
• Beyond Sally-Anne: Evaluating Theory of Mind in LLMs using Epistemic Schelling Points — The Epistemic Asymmetry Schelling Task (EAST) reveals LLMs' significant functional social reasoning and epistemic tracking gaps.
• Embodied Multi-Agent Coordination by Aligning World Models Through Dialogue — Dialogue in embodied LLM agents reduces action conflicts but can degrade task success without genuine world-model alignment.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.