An AI startup has emerged that aims to model changes in human behavior over the long term without relying on role-playing by large language models (LLMs).

TechCrunch reported on Oct. 6 that Mirror Particle in San Francisco is developing its own foundation model to analyze what consumers do and why they behave that way. The company describes it as a “world model” that simulates human motivations and behavior that change over time.

In the behavior prediction market, a widely used approach is to assign LLMs the role of a specific age group or consumer segment, or to create virtual consumers by further training them with additional data. Abhivyakti Ahuja (아비비약티 아후자), co-founder and CEO of Mirror Particle, argued that there are limits to modifying LLMs that have already learned vast amounts of text with small amounts of data, and that it is difficult to model visual and spatial cognition and social intelligence sufficiently with language alone.

Mirror Particle combines client consumer data with current affairs, popular culture and social media, and puts more weight on people’s actual behavior than on survey responses. The key is to analyze a group not as fixed demographics but as a system in which motivations and choices change depending on experiences and events.

Its initial business areas are market research and brand and product strategy. Rather than stopping at producing advertising copy for Gen Z, it analyzes whether that consumer group wants the product itself. In a pilot with a pet food company, it presented a result that the perception of being a “popular and affordable brand” was a bigger reason for stagnant sales than packaging images. This, however, is a case disclosed by the company and is not an independently verified result on predictive accuracy.

Investment is also pouring into the human behavior modeling field. Simile recently raised $200 million, valuing it at $2 billion, and Humans& also raised $480 million in seed funding this year.

Mirror Particle has completed angel funding and is nearing the end of its first venture funding round. Over the long term, it aims to expand into a general-purpose layer that predicts individual behavior beyond group analysis.