String Seed of Thought: Prompting LLMs for Distribution-Faithful and Diverse Generation

Ask a large language model to flip a fair coin a hundred times and something strange happens: the ratio of heads to tails drifts noticeably away from 50:50. The model knows what a fair coin should do in theory, but knowing the target distribution and actually producing outputs that match it turn out to be two very different problems.

Sakana AI says this isn't just a coin-flip curiosity. The same skew shows up when you ask a model to brainstorm story ideas or generate multiple solutions to a problem — the outputs bunch up around a narrow set of options instead of spreading out, which is exactly the opposite of what you want from a brainstorming tool or a recommendation system.

Their fix, called String Seed of Thought (SSoT), is almost embarrassingly simple. You just tell the model to generate a random string as part of its own output, then have it manipulate that string to derive its final answer. No external random number generator, no extra tooling — just a small addition to the prompt itself.

According to Sakana AI, SSoT cuts output bias across a wide range of models, both open and closed. Reasoning models like DeepSeek-R1 get close to the accuracy you'd see from genuine random sampling. The trick isn't limited to binary choices either — it scales up to n-way selections and arbitrary probability distributions. On the NoveltyBench diversity benchmark, SSoT beat other approaches across all six categories tested, and it did so without sacrificing output quality.

The paper is set to be presented at ICLR2026, and Sakana AI frames it as foundational groundwork for anything that needs LLMs to behave probabilistically in the real world — content generation, idea generation, and new approaches to inference-time scaling among them.