LLMs Keep Writing the Same Metaphors. A Startup Is Trying to Fix That.
Ask 25 language models to write a metaphor about time. Most will say time is a river. Some will call it a weaver. Researchers tested exactly this, publishing a November 2024 paper called "Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond)." It won best paper at NeurIPS.
What the Research Found
The study ran 25 LLMs through 50 prompts each, covering top US commercial models and open-source Chinese models. Of 1,250 responses, most converged on similar metaphors. Different companies, different training data, same rivers and weavers.
The obvious workaround is raising temperature, the parameter that controls output randomness. The problem: dialing it to maximum on OpenAI models caused responses to switch from English to code mid-sentence. More randomness globally produces incoherence, not variety.
What Springboards Built
Australian startup Springboards took a more targeted approach. Their model Flint, built on Alibaba's open-source Qwen 3, does not add randomness uniformly. It identifies specific positions in an output where variety is appropriate, then inserts randomness only at those points. The rest stays coherent.
Flint sits alongside ChatGPT and Claude as an alternative model option in the Springboards tool. The company describes it as a prototype aimed at advertising and marketing professionals.
Where This Stands
The homogeneity problem is documented. A NeurIPS best paper is about as validated as AI research gets. The targeted-injection approach is at least theoretically coherent: if global randomness causes incoherence, local randomness at the right token positions might not.
Whether it generalizes beyond ad copy is an open question. Advertising and marketing is a reasonable testbed. Users in that space care about output variety in ways that make failure obvious quickly.
One possibility is that "varied" and "better" end up being the same thing in creative work but diverge in technical contexts. Someone will write that NeurIPS paper eventually.
Source: Technologyreview