String Seed-of-Thought Enhances LLM Probabilistic Option Selection

Researchers introduced String Seed-of-Thought (SSoT), a prompting technique that materially improves Probabilistic Instruction Following (PIF) for large language models. The method asks the model to generate a random string and then extract randomness from that string to deterministically select from a predefined set, which yields much better empirical alignment with target option distributions and larger response diversity. Experiments in the paper show SSoT approaches the behavior of a pseudo-random generator on closed-set tasks and increases diversity on open-ended tasks, evaluated on benchmarks including NoveltyBench.

The core idea of SSoT is to create in-prompt entropy and isolate randomness from semantic decoding. In practice the prompt instructs the model to:

• •output a random alphanumeric string (the "seed"),
• •perform a fixed manipulation on the seed (hash-like transform, modulus mapping, or character-to-index mapping),
• •map the transformed value to the target option set, producing the final answer.

This two-stage flow separates the randomness source from the semantic answer generation and reduces the model tendency to optimize for high-probability tokens that cause option collapse.

LLMs are trained to maximize likelihood and thus prefer high-probability continuations, which makes naive requests for random choices biased. SSoT forces the model to produce low-level entropy first, and then follow deterministic extraction rules to produce an option. That construction turns the model into an approximate pseudo-random sampler without changing sampling hyperparameters like temperature or requiring model fine-tuning.

The paper reports that SSoT substantially narrows the gap to ideal distributions on closed-set tasks (for example, a target 50/50 coin distribution) and avoids the collapse seen in baseline and simple prompt variants. On open-ended generation, SSoT increases measured diversity on NoveltyBench by steering outputs away from repeated high-probability modes. The authors test multiple mapping strategies and show robustness across seeds and prompt phrasings.

SSoT is model-agnostic and zero-shot at inference time, so it is an immediate tool for systems that need non-deterministic behavior: simulated user behavior, multi-agent game-playing, randomized A/B creative generation, and stress-testing. It can complement sampling strategies (--temperature, nucleus sampling) by providing structural randomness rather than relying solely on token-level stochasticity. Implementers should validate mapping functions to avoid accidental biases (character sets, length constraints) and guardrails if seeds could leak sensitive structure.

Prompt engineering has accumulated many pattern-based fixes for model idiosyncrasies. SSoT addresses a persistent gap: LLMs are strong at deterministic reasoning but weak at matching explicit probability distributions across repeated draws. The method sits alongside other prompting strategies like chain-of-thought for reasoning and self-consistency for robustness, but is unique in turning prompted outputs into a controlled pseudo-random process. It also speaks to a broader theme: behavioral properties often need architectural or procedural scaffolding around probability, not just raw model capacity.

Validate SSoT across commercial API models and edge cases where seed-to-choice mapping could interact with safety filters or tokenization. Expect community variants: cryptographic-hash style mappings, external PRNG seeding coupled to model prompts, and integrations that combine SSoT with reinforcement learning or multi-agent coordination. The next questions are reproducibility at scale and interactions with safety and content policies.