New 'Seed-of-Thought' Prompting Technique Solves AI's Core Randomness Problem

The fundamental paradox of modern large language models is that they appear endlessly creative, yet they are fundamentally deterministic engines trapped by the statistical gravity of their own training data. When a user interacts with an AI, the model is not truly "thinking" in a spontaneous manner; it is calculating the most probable sequence of words to follow the prompt. While this architecture makes AI exceptionally good at retrieving facts and writing standard code, it creates a severe bottleneck when a task requires genuine unpredictability or out-of-the-box thinking.[6]

When asked to brainstorm ten different marketing campaigns or write a novel story, an LLM will almost always gravitate toward the statistical mean of human thought. This phenomenon, known in the industry as "mode collapse" or "typicality bias," makes AI outputs increasingly predictable and repetitive. As models become smarter and more heavily aligned to be safe and helpful, they simultaneously lose their creative edge, resulting in a landscape where every AI-generated response begins to sound exactly the same.[2][4]

The root of this predictability lies in a technical limitation known as Probabilistic Instruction Following (PIF). As researchers have long noted, large language models are exceptionally bad at generating true randomness or selecting from a set of options based on a specific target probability. Because they are designed to find the single "best" answer, they struggle to understand that sometimes the correct answer is to be intentionally arbitrary.[3][4]

Consider a simple test: asking an AI to simulate flipping a fair coin 100 times. Instead of returning a 50/50 split of heads and tails, the model's output will skew wildly. This drift occurs because the AI is influenced by hidden biases, such as the position of the words in the prompt or the frequency of the word "heads" in its vast training corpus. The AI understands the concept of a 50/50 probability, but its underlying architecture prevents it from faithfully executing it.[1][3]

This inability to follow probabilistic instructions has crippled applications that rely on non-deterministic behavior. For developers trying to simulate human interactions in game theory, build unpredictable non-player characters in video games, or generate diverse synthetic data to train next-generation models, the AI's inherent bias is a massive hurdle. Until recently, the only solutions involved relying on external random number generators or writing complex, multi-step code to force the AI out of its rut.[4]

Now, a breakthrough from Tokyo-based research lab Sakana AI offers an elegant, lightweight solution. Researchers Kou Misaki and Takuya Akiba have developed a prompting technique called "String Seed of Thought" (SSoT), which effectively teaches large language models how to roll dice in their own heads. By making a minor adjustment to how the AI is instructed, the researchers have unlocked the ability for models to generate true internal randomness.[3]

Presented in a paper accepted at the prestigious International Conference on Learning Representations (ICLR) 2026, the SSoT technique solves the randomness problem without requiring complex API integrations, external tools, or expensive model retraining. It is a pure prompt engineering solution that works across a wide variety of both open-source and closed-source models, making it immediately accessible to anyone interacting with an AI.[5]

The technique relies on a deceptively simple two-stage prompt. First, the user instructs the large language model to generate a complex, random string of characters—for example, an 8-character alphanumeric sequence—directly in its own output window. The prompt explicitly tells the AI to make this string appear sufficiently complex and unpredictable, with no obvious structure or pattern.[1][3]

Second, the prompt instructs the model to use that newly generated string as a "seed" to shape its final answer. If the task is to flip a coin, the AI mathematically manipulates the random string to sample from the target 50/50 distribution. If the task is creative writing, the AI uses the string's entropy to force itself down a less probable, more unique narrative pathway.[1][4]

The brilliance of this approach lies in how it bypasses the model's inherent training biases. When an LLM tries to answer a question directly, it immediately falls into the gravitational pull of its training data's most common answers. However, generating a random string is a task-agnostic action with no inherent semantic bias. The model has no preconceived notion of what an 8-character string "should" mean.[4]

By forcing the model to generate this string first, SSoT injects what researchers call "structured entropy" into the model's context window. The LLM then uses its own advanced reasoning capabilities to map that entropy to the desired output. Because the seed is different every time, the resulting output is forced to be different every time, effectively breaking the deterministic chain.[2][4]

The empirical results of the SSoT technique are striking. When tested on advanced reasoning models like DeepSeek-R1, the technique allowed the AI to achieve sampling faithfulness that approaches the performance of an actual pseudo-random number generator (PRNG). For the first time, an LLM could reliably flip a coin in its head and land on a perfect 50/50 distribution over thousands of iterations.[3]

Beyond simple probability tests, SSoT dramatically improves open-ended creative tasks. On the NoveltyBench diversity benchmark, the technique outperformed all other prompting approaches across six different categories. By using the random seed, the models were able to explore the outer edges of their latent space, generating highly diverse story ideas, marketing concepts, and brainstorming solutions while maintaining high output quality.[3][4]

This breakthrough arrives at a critical moment in artificial intelligence development, addressing what industry analysts have dubbed the "Alignment-Diversity Paradox." As AI companies race to make their models safer and more reliable, they inadvertently make them more boring.[2]

Rigorous safety training methods, such as Reinforcement Learning from Human Feedback (RLHF), train models to behave in ways that human reviewers prefer—polite, factual, and consistent. However, studies show that as models become heavily aligned, they lose their epistemic diversity. SSoT offers a way to reclaim that lost creativity and randomness without compromising the model's underlying safety guardrails.[2]

For prompt engineers and everyday creators, SSoT acts as a lightweight revolution. Instead of endlessly tweaking prompts to beg the AI for "more unique" or "less cliché" ideas—which often just results in a different flavor of cliché—users can simply inject the SSoT instruction to guarantee a wider, mathematically sound spread of concepts.[2]

The implications extend deeply into social science and economics. Researchers are increasingly using large language models as proxies for human participants in game theory simulations and behavioral studies. SSoT ensures that these AI agents can employ "mixed strategies"—making unpredictable, probabilistically sound choices—just as real humans do, making synthetic research far more reliable.[4]

Ultimately, the String Seed of Thought technique proves that the perceived limitations of large language models are often just failures of human instruction. By understanding how these neural networks process entropy, developers have unlocked a new layer of capability, proving that even the most deterministic machines can be taught to embrace the unpredictable.[6]