Mitigating Social Desirability Bias in Random Silicon Sampling

Name: Mitigating Social Desirability Bias in Random Silicon Sampling
Uploaded: 2026-10-10T12:37:39.739259+00:00
Duration: 626 s
Description: Researchers used Large Language Models (LLMs) to simulate large populations for surveys by conditioning them on demographic data. They tested five prompting techniques to reduce 'Social Desirability Bias' (SDB), which is the tendency for LLMs to give socially approved answers. The study found that question reformulation was the most effective method for improving alignment and response diversity.

Researchers used Large Language Models (LLMs) to simulate large populations for surveys by conditioning them on demographic data. They tested five prompting techniques to reduce 'Social Desirability Bias' (SDB), which is the tendency for LLMs to give socially approved answers. The study found that question reformulation was the most effective method for improving alignment and response diversity.

Open the full episode