String Seed-of-Thought Enhances LLM Probabilistic Option Selection
String Seed-of-Thought (SSoT) is a new prompt engineering method designed to make large language models follow probabilistic instructions more faithfully. SSoT asks the model to first emit a random string, then deterministically map that string to a choice, which injects entropy and reduces the output collapse that causes biased, non-diverse selections. The technique improves closed-set probabilistic tasks, like coin flips or multi-option selection, and boosts open-ended diversity in benchmarks such as NoveltyBench. SSoT is lightweight, model-agnostic, and works at prompt time without fine-tuning or sampling tweaks, making it immediately useful for simulation, game AI, content diversification, and any application that needs distributional fidelity.
