Large language models (LLMs) are increasingly powering interactive agents in virtual worlds, from simulated towns to online games. But one challenge remains: these agents often produce repetitive, predictable conversations. New research from the University of Tokyo explores how to make these AI-driven chats more diverse and engaging. Researchers found that by strategically pruning the information given to the LLM before it generates a response, the agents could have more creative and varied conversations. Think of it like limiting an author's access to a thesaurus – fewer readily available words might force them to come up with more novel phrasing. This 'Adaptive Prompt Pruning' method selectively removes parts of the input prompt, like memories or previous dialogue, based on what the model focuses on most (its 'attention weights'). By removing these highly attended-to pieces, the LLM is nudged to think outside the box. The study also found that the order and length of information within the prompt impact how diverse the generated responses are. Longer prompts, while providing more context, can actually stifle creativity by over-constraining the model. Finally, the researchers addressed a crucial challenge: ensuring that increased diversity doesn’t lead to agents contradicting themselves. They added a “correction step” after the response is generated to reconcile any inconsistencies with the pruned information, ensuring that while agents are more creative, they still stay true to their established backstories and previous conversations. This research opens up exciting possibilities for more dynamic and realistic AI-powered simulations. Imagine virtual worlds populated by agents who can surprise you with their unique perspectives and unpredictable interactions. Overcoming the current limitations of LLM-driven conversations could lead to more engaging games, more realistic simulations, and a deeper understanding of how humans communicate.