How Can LLMs Learn to Mimic Human Preferences? From Synthetic Consumers to World Simulation

Session Overview

Simply assigning an LLM a persona based on demographic probabilities is not enough to reproduce how real consumers respond. Intellisia initially developed Asia’s first synthetic consumer technology using a demographic approach, but complex enterprise surveys produced a reproduction accuracy of only 67%.

To overcome this limitation, Intellisia developed a Reverse Refinery method that reconstructs consumer profiles from hundreds of thousands of real responses. By modeling more natural, human-like response distributions and applying fine-tuning to correct the ethical guardrails built into LLMs, the company increased synthetic consumer reproduction accuracy to 91.2% within one year.

This session presents the research methodology and lessons learned from numerous failed experiments, along with insights gained from 120 proof-of-concept projects with 60 companies. It also explores the next step beyond individual consumer simulation: building a World Simulation Company capable of modeling entire markets and societies.

Key Takeaways

• Why simple demographic personas fail to reproduce real consumer behavior
• How the Reverse Refinery method reconstructs consumers from real response data
• Using fine-tuning and ethical guardrail calibration to create more human-like response patterns
• How synthetic consumer technology achieved 91.2% reproduction accuracy across 120 PoCs with 60 companies
• The future of World Simulation beyond individuals, extending to markets and society