"Don't Say It!": Constraints, Compliance, and Communication when Language Models Play Taboo

2026-07-01

AI Summary

This study investigates how large language models (LLMs) balance rule adherence and communicative effectiveness under strict lexical constraints, as exemplified by the “Taboo” word game. By systematically intervening at multiple levels of the generation pipeline—including prompt engineering, constrained decoding, and manipulation of internal representations—and integrating taboo-word detection, LLM-as-a-judge evaluation, and human guesser experiments, the work presents the first joint assessment of rule compliance, communication efficacy, and alignment with human strategic behavior in constrained language generation. The findings reveal significant trade-offs between different intervention strategies in satisfying lexical restrictions while maintaining expressive power. Moreover, current LLMs perform substantially worse than humans when acting as guessers, highlighting that grounding language in contextually relevant lexical knowledge remains a critical challenge in constrained communication settings.

Abstract

The game of Taboo requires describing a target word without using a set of forbidden words, so that other players can guess it. This deceptively simple task combines strict lexical constraints with the need for communicatively effective descriptions, making it a compelling playground for examining how LLMs navigate competing demands at inference time. We evaluate two open-weight models under conditions that intervene at progressively deeper levels of the generative process, from prompting to generation-time constraints to internal representations manipulations. We assess their outputs through forbidden word violation detection, LLM-as-a-judge measuring the degree to which generated descriptions successfully evoke the target concept for both human and machine guessers, and examining whether the strategies models adopt under constraint align with those of human players. Our results show that compliance with the rules of the game and communicative effectiveness trade off differently across conditions, and that models remain substantially weaker than humans as guessers, suggesting that lexical grounding under constraint is an open challenge for current language models.

Problem

• lexical constraints
• communicative effectiveness
• language models
• Taboo game
• lexical grounding

Innovation

• lexical constraints
• communicative effectiveness
• generation-time intervention
• lexical grounding
• LLM-as-a-judge