Sakana AI Highlights New Technique to Reduce Bias in Large Language Model Outputs

According to a recent LinkedIn post from Sakana AI, the company is highlighting research on a prompting method designed to reduce bias in large language model (LLM) outputs. The post describes how standard LLMs tend to deviate from expected probability distributions in simple tasks such as simulated coin flips and in more complex creative-generation scenarios.

The LinkedIn post explains that Sakana AI’s team has developed a technique called String Seed of Thought (SSoT), which asks an LLM to generate a random string and then uses that string to derive an answer. According to the post, this method mitigates output bias across a range of open and closed LLMs and reportedly approaches the accuracy of true random sampling when used with reasoning-focused models.

The post further suggests that SSoT generalizes from binary decisions to multi-way choices and arbitrary probability distributions, and that it improves diversity scores on the NoveltyBench benchmark while maintaining output quality. The work is slated for presentation at ICLR 2026, signaling some degree of peer recognition within the AI research community.

For investors, this research focus may indicate Sakana AI’s intention to differentiate its technology in areas where probabilistic fidelity and diversity of generation are commercially important, such as recommendation systems, brainstorming tools, and creative content platforms. If the technique is adopted broadly or embedded into products, it could enhance the company’s competitive position in the LLM tooling and infrastructure market, though the post does not provide information on commercialization timelines or revenue impact.