"Don't Say It!": Constraints, Compliance, and Communication when Language Models Play Taboo
We use the game of Taboo to study how LLMs handle strict lexical constraints while still trying to communicate. Comparing prompting, constrained decoding and SAE-based steering, and grounding the evaluation in a human study, we find that rule compliance and description quality trade off differently depending on how the constraint is imposed, and that models remain far weaker than humans at guessing.
Jul 1, 2026
