This AI model can make creative connections puzzles; Inspired by The New York Times' hit, researchers put an LLM to the task

Timothy Merino (left) and Julian Togelius (right) map ideas for their Connections research on a whiteboard.

Timothy Merino (left) and Julian Togelius (right) map ideas for their Connections research on a whiteboard. 

NYU researchers, led by Associate Professor Julian Togelius and PhD student Timothy Merino, have successfully used OpenAI's GPT-4 to create puzzles for the popular game Connections. Their study found that while large language models (LLMs) struggle with metacognition, they can produce engaging puzzles when given specific prompts and subtasks. Merino noted LLMs' vast knowledge base as an advantage, while Togelius suggested the method could be applied to other word games. The team's approach resulted in AI-generated puzzles comparable to those created by The New York Times, as confirmed by human player evaluations.