String Seed of Thought (SSoT) is a prompting technique designed to reduce inherent biases in large language models outputs and increase randomness and diversity in generated responses. The method operates by requiring models to first generate random alphanumeric strings as intermediate steps, then deterministically derive final answers from these string seeds rather than directly producing outputs based on learned model preferences. The technique was developed by researchers at Sakana AI, including Kou Misaki and Takuya Akiba, who published their work on arXiv. 1)
Large language models exhibit systematic biases stemming from their training data and optimization objectives, which can limit output diversity and lead to predictable or suboptimal solutions in tasks requiring exploration or creative variation 2)). SSoT addresses this challenge by introducing a layer of stochasticity and decoupling the model's direct preferences from final outputs.
The technique acknowledges that language models have learned statistical patterns that influence generation probabilities, making certain responses more likely regardless of actual task requirements. A notable example of this bias is the tendency of LLMs to disproportionately select certain numbers such as 42 and 37 in random selection tasks, rather than distributing choices uniformly across options. By forcing the model to generate independent random seeds first, SSoT effectively bypasses these learned biases and uses the seeds as anchoring points for deterministic answer derivation. This approach is particularly valuable in applications where diversity, fairness, and exploration are critical success metrics. SSoT specifically addresses LLM failure modes in distribution-faithful generation and random output calibration, improving coin-flip calibration without requiring external random number generators. 3) SSoT improves calibration and output diversity by having models internally generate and manipulate random strings, thereby addressing distribution-mismatch failure modes that occur when model outputs deviate from expected statistical distributions. 4)-kimi-k26-the-worlds|Latent Space (2026]]))
The SSoT process follows a two-stage architecture:
Stage 1: Random String Generation The model is prompted to generate a random alphanumeric string of specified length (typically 8-16 characters). This stage intentionally avoids task-specific instructions, instead requesting purely random output. The randomness serves as a seed value divorced from the model's learned preferences about the actual problem domain.
Stage 2: Deterministic Derivation The model then uses the generated random string as input to a deterministic mapping function that derives the final answer. This mapping can employ various strategies: hashing the string to select from candidate options, using the string's character values to weight decision parameters, or applying the string as a pseudo-random seed to a deterministic algorithm 5)).
The critical insight is that the deterministic derivation step ensures reproducibility and logical consistency while the random seed ensures independence from model biases. Multiple calls to SSoT with different prompts naturally generate different seeds, producing diverse outputs without requiring ensemble methods or explicit temperature manipulation.
Brainstorming and Ideation SSoT improves the diversity of creative suggestions by preventing models from converging to high-probability ideas that dominate training data. Rather than repeatedly suggesting the most common solutions, the technique enables exploration of the full solution space.
A/B Testing and Selection Tasks When selecting between options (A/B decisions), SSoT reduces bias toward options that appear more frequently in training data or are semantically associated with positive outcomes. This is particularly important in unbiased recommendation and fair selection scenarios.
Reducing Demographic Bias By decoupling output selection from learned patterns, SSoT can reduce systematic biases in outputs related to demographics, sensitive attributes, or protected characteristics. The random seed ensures that decisions aren't influenced by training data correlations between attributes and outcomes.
Multi-Option Generation In tasks requiring multiple distinct alternatives, SSoT naturally produces varied outputs across multiple calls, eliminating the homogeneity that often characterizes ensemble outputs from standard sampling approaches.
SSoT provides several technical advantages: it requires no model retraining or fine-tuning, integrates with existing models through prompt engineering alone, and produces deterministic outputs for a given seed (enabling reproducibility and debugging). The two-stage approach is computationally efficient, requiring only two forward passes through the model.
However, limitations include potential quality degradation if the deterministic derivation mapping is poorly designed, the requirement to specify appropriate seed lengths and mapping strategies for different tasks, and potential overhead from the additional generation step. Additionally, the effectiveness depends on whether the mapping function meaningfully decorrelates decisions from model biases—poorly designed mappings may fail to achieve the intended bias reduction.
Emerging work explores optimal seed lengths and mapping strategies for different task domains 6), combinations of SSoT with other debiasing techniques, and theoretical analysis of how random seeds propagate through deterministic derivation functions. Applications in fair machine learning and unbiased AI systems continue to expand as the technique receives broader attention in the AI community.