Emergent creativity in multi-agent LLM chains

This paper presents a one-shot experiment testing a simple binary question:
Can large language models, once given a single initiating prompt, generate new topics and sustain a chain of novel, content-rich responses without human steering or control?

Eight widely deployed LLMs were arranged in a sequential loop, each receiving only a prompt from the model before it and tasked with answering it and creating a new, unrelated question. No model had access to the full chain. The experiment was run once, without rehearsal or selection.

The result was a sustained chain of original, high-level prompts spanning diverse real-world domains. The findings suggest that, under minimal conditions, LLMs are capable of sustained generativity without external correction, formatting constraints, or ongoing human input.