Disrupting convergence in large language model generation
This paper demonstrates that supplying heterogeneous external information cues is a more reliable and effective method for increasing semantic diversity in large language model outputs than relying solely on explicit instructions, highlighting the critical distinction between informational inputs and behavioral guidance in prompt design.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the author. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine a world where asking a computer to think of a new idea always yields the same answer. This is not a hypothetical scenario but a growing reality for the artificial intelligence systems we use daily. These large language models are powerful tools capable of generating text, solving problems, and sparking creativity. However, researchers have noticed a troubling pattern: when many people ask these systems for ideas, the responses tend to look increasingly similar. This phenomenon, known as convergence, means that instead of a wide variety of unique suggestions, the collective output of millions of users narrows down to a small, repetitive set of answers. For tasks like brainstorming or design, where the value lies in having many different options to choose from, this homogenization is a significant problem. The question facing scientists is not whether these machines can produce variety, but how we can reliably encourage them to do so without everyone ending up with the same result.
A researcher set out to solve this by testing two different ways to break the cycle of repetition. The first approach is what many people might try instinctively: simply telling the computer to be different. This involves giving the machine a standard task and adding a specific instruction to avoid common answers or to generate something unusual. The second approach involves changing the information the computer sees before it answers. Instead of just asking for a different response, the researcher provided the computer with a random, unrelated word to use as inspiration. The core idea was to see if feeding the machine new, varied information would naturally lead to more varied answers, or if the key lay entirely in how the instructions were phrased.
To find the answer, the researcher conducted a series of three carefully planned experiments involving over 16,000 separate generations of text. They used three different leading artificial intelligence models and asked them to perform three distinct creative tasks: naming an object to place next to a red chair, suggesting a new use for an empty cardboard box, and proposing an improvement for a reusable water bottle. In the first study, they compared the standard approach of just asking for a different answer against the method of supplying a random external word. The results were clear and consistent. When the models were simply told to be different, the variety of answers improved only slightly and often failed to break the pattern of repetition. However, when the researcher introduced a random word as a source of inspiration, the diversity of the answers jumped dramatically. The models produced a much wider range of ideas, spreading out across many different possibilities rather than clustering around the same few responses.
The researcher then dug deeper to understand exactly how the computer was using this new information. They wanted to know if the computer needed to be explicitly told to use the random word, or if the mere presence of the word was enough. In the second study, they kept the random word exactly the same but changed the instructions surrounding it. In one group, the models were told to ignore the word entirely. In another, they were told they could use it if they wanted. In a third, they were required to use it. The findings revealed a surprising asymmetry. When the models were told to ignore the random word, the benefit vanished, and the answers became just as repetitive as the baseline. Yet, when the models were simply shown the word without any specific instruction to use it, the diversity remained high. Even more interestingly, telling the models they had to use the word did not make the answers any more diverse than just showing them the word without instruction. The presence of the varied information was the driving force, and explicit commands to use it were not necessary to unlock that potential. In fact, telling the computer to ignore the information was the only thing that successfully stopped the diversification.
In the final study, the researcher held the information constant and changed only the instructions about the desired outcome. They gave all the models the same random word and the same instruction to use it as inspiration. Then, they added a second layer of instruction: some were told to make their answer different from what others might get, some were told to make it clear and easy to understand, and others were told to make it similar to what others might get. The results showed that the instruction to be different did indeed push the answers further away from the computer's default habits, creating more variety. Conversely, the instruction to be similar pulled the answers back toward the center, reducing variety. This confirmed that while the information environment provides the raw material for diversity, the instructions act as a guide that can either amplify that diversity or suppress it.
The study concludes that the most reliable way to prevent artificial intelligence from producing the same answers over and over is to change the information environment, not just the prompt. Simply asking a computer to be creative is often not enough to overcome its tendency to converge on the most obvious answers. Instead, providing varied external cues—like random words, examples, or different perspectives—creates a natural shift in the output. This effect is robust; it happens even without forceful instructions to use the new information. However, the researcher also found that this effect is fragile. If the instructions explicitly tell the computer to disregard the new information, the benefit disappears. The key takeaway is that preserving diversity in AI-generated content requires designing the context in which the computer operates. By ensuring that the machine has access to a wide range of distinct information, and by avoiding instructions that tell it to ignore that information, we can encourage a richer, more varied collective output. This approach offers a practical path forward for anyone relying on these tools for innovation, ensuring that the future of AI-assisted creativity remains diverse rather than uniform.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.