Don’t Say It! Constraints, Compliance, and Communication when Language Models Play Taboo

Jul 16, 2026

Upload by Patrick

The game of Taboo requires describing a target word without using forbidden terms, combining strict lexical constraints with the demand for communicatively effective descriptions so the guesser can identify the target. This study evaluates two open-weight language models as Taboo players under interventions at progressively deeper generation stages: prompting, generation-time constraint enforcement, and manipulations of internal representations. Outputs are measured via forbidden-word violation detection, LLM-as-a-judge concept evocation, and comparisons with human strategies. Results show a condition-dependent trade-off between rule compliance and communicative effectiveness, and overall weaker performance than humans in the guessing role, indicating unresolved lexical grounding under constraints.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages21

Reading Time53 min

Abstract

Introduction

Related works

Language games as NLP benchmarks

Constrained text generation

"Don’t Say It!": Constraints, Compliance, and Communication when Language Models Play Taboo Sara Candussio1,†, Francesca Padovani2,†, Daniel Scalena2,3,† and Malvina Nissim2 1 AILab, MIGe, Univeristy of Trieste, Italy 2 Center for Language and Cognition (CLCG), University of Groningen, The Netherlands 3 University of Milano-Bicocca, Italy 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 models1 . Keywords taboo, prompting, constrained generation, word-guessing, SAEs 1. Introduction The game of Taboo requires a player to describe a target word without using a set of forbidden words, so that another player can guess it. Beyond its appeal as a parlour game, Taboo instantiatesa linguistically interesting challenge: the speaker must suppress some lexically and conceptually salient words while simultaneously producing a description that is informative enough to identify the target. This combination of constraint satisfaction and communicative effectiveness makes it a particularly suitable setting for probing language models in a setting that goes beyond standard instruction-following benchmarks. In particular, it remains unclear whether descriptions generated under constraint conform to the demands ofthe task — namely whether 1Code and data can be found at [https://github.com/DanielSc4/LMtaboo/](https://github.com/DanielSc4/LMtaboo/) CLiC-it 2026: Twelfth Italian Conference on Computational Linguistics, September 14 — 16, 2026, Palermo, Italy * Corresponding author. † These authors contributed equally. $ sara.candussio@phd.units.it (S. Candussio); [f.padovani@rug.nl](f.padovani@rug.nl) (F. Padovani); [d.scalena@rug.nl](d.scalena@rug.nl) (D. Scalena); [m.nissim@rug.nl](m.nissim@rug.nl) (M. Nissim) 􀀚 0009-0004-5198-6970 (S. Candussio); 0009-0007-3489-9631 (F. Padovani); 0009-0006-0518-6504 (D. Scalena); 0000-0001-5289-0971 (M. Nissim) © 2026 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0) . they are sufficiently detailed, salient enough to evoke the target concept, and communicatively effective enough for the game to be successfully completed. In this paper we investigate how LLMs play Taboo in Italian, evaluating two open-weight models under conditions that intervene at progressively deeper levels of the generative process: from prompting, to generation-time constraints, to the manipulation of internal representations. To ground our evaluation in an actual game-play, we conduct a human study in which the roles are reversed in both directions: human evaluators read model-generated descriptions and attempt to guess the target word, and vice versa. This bidirectional design allows us to directly assess how well descriptions produced succeed in conveying the target concept, and whether the descriptive cho

FAQs

What is the Taboo task used to evaluate language models?

In Taboo, a player must describe a target word without using a predefined set of forbidden words so that another player can guess the target.

How does the paper test models at different levels of the generation process?

It evaluates two open-weight models under interventions that progress from prompting, to generation-time constraints, to manipulations of internal representations.

How are model outputs assessed for both rule-following and effectiveness?

Assessment combines forbidden word violation detection, an LLM-as-a-judge measuring how well descriptions evoke the target concept for human and machine guessers, and analysis of how adopted strategies compare to human players.

Jul 16, 2026

Upload by Patrick