This study examines how large language models navigate the competing demands of strict lexical constraints and communicative effectiveness by playing the game of Taboo. The researchers evaluated two open-weight models under conditions that intervene at progressively deeper levels of the generative process, ranging from prompting to internal representation manipulations.
- Outputs were assessed through forbidden word violation detection and LLM-as-a-judge metrics measuring how well descriptions evoke the target concept for human and machine guessers.
- The analysis compared the strategies adopted by models under constraint against those used by human players.
- Results indicate that compliance with rules and communicative effectiveness trade off differently across conditions.
The findings suggest that lexical grounding under constraint remains an open challenge, as models remain substantially weaker than humans in this task.