Mirror Particle, a two-year-old San Francisco startup, is building from scratch a foundational "world model" that predicts human behavior and models how it changes over time. The company's founder and CEO, Abhivyakti Ahuja, believes that prompting language models to play demographic roles or fine-tuning them for that purpose is a fundamentally wrong approach. TechCrunch reported this on October 6, 2026. The company said it has already closed its angel round and is close to closing its first venture round.
Why the Company Rejects LLM Role-Play
Mirror Particle's central thesis is that forcing a large language model to "play" a member of a specific demographic group is not a reliable way to predict human behavior. This approach relies on the model's language abilities, but its responses reflect text statistics, not real behavioral patterns. That is why the company is building a dedicated "world model" constructed from scratch on behavioral data rather than language — designed to predict what people do and how their behavior changes over time.
"Prompting language models to play demographic roles or fine-tuning them for that purpose is a fundamentally wrong approach," Mirror Particle's leadership said in a comment to TechCrunch.
Telling a language model to answer as a representative of a certain age, income, or region has become standard practice in behavior prediction. Mirror Particle is questioning the reliability of this practice: in its assessment, such "role-play" reproduces not the decision-making logic of real demographic groups, but stereotypes from internet texts. Instead, the company proposes a model that learns from observable signals — people's purchases, the content they watch, and social activity.
Independent reviewers note that this central claim remains the company's own thesis for now: no published benchmarks or independent verifications showing that a purpose-built behavioral model outperforms LLM role-play have been disclosed.
"Revealed Behavior": Real Actions, Not Surveys
Mirror Particle combines data its clients hold about their customers with current events, pop culture, and social media signals. The key to the approach is "revealed behavior" — not what people say about themselves in surveys, but what they actually do. The company leans on exactly this distinction: analyzing real behavioral traces instead of self-reporting errors and socially expected answers yields more accurate predictions, it argues.
Surveys have been the primary tool of market research for decades, but their flaws are well known: respondents misremember their behavior or give socially acceptable answers. The "revealed behavior" approach sidesteps this problem — it measures not what people say, but what they do. Mirror Particle plans to scale this principle to the level of a large-scale model: when client data merges with external signals, brands will be able to anticipate audiences' future actions, the company argues.
The initial go-to-market strategy targets market research and brand and product strategy. In other words, the company sees brands and research agencies that need to understand consumer behavior as its first customers. Traditional survey methods still dominate this segment, and Mirror Particle aims to replace them.
The Competitive Field Is Getting Crowded
Competition in human behavior modeling has intensified sharply over the past year. Simile raised $200 million at a $2 billion valuation, and Aaru raised $88 million at a $1 billion valuation. Humans& announced a $480 million seed round at a $4.48 billion valuation while also launching Persimmon, a product designed for modeling human behavior.
Against this backdrop, two-year-old Mirror Particle is still at a relatively early stage: the company said it has closed its angel round and is approaching the close of its first venture round. Details of the round — its size and participants — have not yet been disclosed.
The Team: From Neuroscience to Robotics
Abhivyakti Ahuja studied neuroscience and computer science at the University of Toronto. He was inspired by the research of Geoffrey Hinton, known for his work on neural networks. After graduating, he worked at Amazon Robotics building robots — where he met future co-founders Will Song and Thomson Yen. The team's experience in neuroscience and robotics appears to have directly shaped the company's "world model" concept: both fields rely on observing the environment and predicting behavior.
Next Test: Startup Battlefield 200
On October 13–15, Mirror Particle will compete in Startup Battlefield 200 as part of TechCrunch Disrupt 2026 in San Francisco. The competition stage will give the company a chance to defend its "world model" thesis before investors and industry experts. At the same time, until independent verifications and open benchmarks are published, the company's central claim — the superiority of a purpose-built behavioral model over LLM role-play — remains an unconfirmed thesis.
