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

Researchers specializing in artificial intelligence published a technical paper on the arXiv preprint server on October 28, 2024, introducing 'String Seed of Thought' (SSoT), a novel prompting technique designed to help large language models (LLMs) more accurately follow probabilistic instructions. This development aims to solve a persistent issue where AI systems struggle to produce diverse outputs that match a specific target distribution when asked to choose from multiple options with varying likelihoods. By implementing SSoT, the researchers found that models could better mimic the expected randomness and statistical variety required for sophisticated simulations and data generation tasks. The core of the research focuses on a concept named Probabilistic Instruction Following (PIF). In standard operational modes, LLMs generally excel at deterministic tasks—those with a single correct answer—but often fail when tasked with selecting an answer from a predefined set of options according to specific probability percentages. For example, if a model is told to generate the word 'A' 70% of the time and 'B' 30% of the time, standard prompting often results in an empirical distribution that deviates significantly from the target, as the model's internal sampling mechanisms are not inherently aligned with complex external constraints. The proposed String Seed of Thought method acts as a middle-ware reasoning step that bridges the gap between the instruction and the final generation. By forcing the model to articulate its sampling logic or follow a specific internal 'string' of thought before making a selection, the researchers demonstrated that the resulting output distribution becomes much more faithful to the human-specified parameters. This shift from simple zero-shot prompting to structured probabilistic reasoning represents a significant advancement in the reliability of synthetic data generation and AI-driven behavioral modeling, where diversity and statistical accuracy are paramount.