Evaluating LLM compliance and communication using Taboo constraints
July 2, 2026
This study evaluates how open-weight models navigate competing demands between lexical constraints and communicative effectiveness in a Taboo-style task. Researchers tested interventions ranging from prompting to internal representation manipulation to measure rule compliance and concept evocation.
HOW THIS AFFECTS YOU
●
researcherYou can use this framework to study how models balance instruction following with semantic utility.
●
policyThis highlights how models behave when forced to operate under strict, conflicting linguistic constraints.