Mirror Particle Wants to Replace LLM Role-Play With a Real World Model of Human Behavior

Synthetic human behavior prediction is suddenly a crowded, well-funded space. Simile raised $200 million at a $2 billion valuation. Aaru pulled in $88 million at a $1 billion valuation. Humans&, founded by alums of Anthropic, xAI, and Google, announced a $480 million seed round in January at a $4.48 billion valuation and launched a product called Persimmon to model human behavior. The money is clearly flowing toward one core bet: that AI can simulate what people will do before they do it.

Most of these approaches lean on the same foundation, large language models prompted or fine-tuned to role-play as a target demographic. Ask the model to "act like a 24-year-old Gen Z shopper" and see what it says about a new skincare line. Mirror Particle, a two-year-old San Francisco startup, thinks that entire premise is broken.

"It's like bringing a super soaker to Niagara Falls," says co-founder and CEO Abhivyakti Ahuja. Her argument is pretty simple when you break it down: LLMs are trained on hundreds of billions of data points of written text, and fine-tuning on a comparatively tiny dataset of customer behavior isn't going to meaningfully shift how that model reasons. "How much can you influence its behavior by fine-tuning with such a small amount of data? It's still stuck in the past."

There's a deeper architectural objection here too. Ahuja doesn't think LLMs model the world the way humans experience it. "LLMs are modeling written language, but humans are made of visual perception, spatial reasoning, social intelligence." Lean on a text-prediction engine to simulate a person, and you inherit all its blind spots. You end up with insights shaped by what humans fail to write down, which is close to useless if your goal is predicting what they'll actually do.

Building a Foundation Model From Scratch

Mirror Particle's bet is a world model, a term borrowed from AI research that generally refers to a system trained to internally simulate how an environment behaves and changes, rather than just predicting the next word in a sequence. Instead of fine-tuning an existing LLM, the company is building its own foundation model designed specifically to simulate why humans behave the way they do, and how that behavior shifts over time.

"We don't want to capture the static person," Ahuja said. "We want to capture the changing person." That means tracking longitudinal data on how people evolve, what triggers move them, and by how much. Stasis counts as data too. "If they aren't changing, that's also a signal."

The model ingests a proprietary mix of inputs: client customer data, current events, pop culture, social media activity, and more. Rather than treating a demographic segment as a fixed profile, Mirror Particle treats it as a system that evolves, tracking how motivations shift as that group moves through real-world experiences. A big piece of the methodology is prioritizing "revealed behavior," what people actually do, over self-reported survey answers, which are notoriously unreliable proxies for real decision-making.

The go-to-market plan looks familiar even if the architecture doesn't. Like its better-funded rivals, Mirror Particle is starting where the budgets already exist: market research and brand strategy. The pitch isn't just better ad copy. It's answering a more fundamental question that brands often get wrong.

Take the beauty industry example Ahuja uses. A brand might ask Mirror Particle to help write Gen Z-targeted copy for an eyeshadow palette. The model's job is to check whether that's even the right product. "What if [the target demographic] doesn't want eyeshadow palettes?" Ahuja said. "Maybe blush is a better option to go for if you want to sell a product to this market." The system doesn't just predict an outcome, it's built to surface the "why" behind it, the motivations and constraints that justify a given recommendation.

One pilot with a well-known pet food brand illustrates the point well. The brand wanted help deciding what imagery to put on packaging, chicken, beef, vegetables, to boost sales. Mirror Particle's model found the imagery question was a dead end entirely. The brand's real problem was that it had become so recognizable it registered as mass-market and cheap in consumers' minds. No amount of packaging tweaks would move sales until that perception issue got addressed directly.

Ahuja frames the model's development trajectory in developmental terms. "The way we see our model evolving is like how a baby learns about the world," she said, pointing to the progression babies go through: vision first, then language, then body awareness, then social intelligence. It's a notably different roadmap than "scale up the transformer and fine-tune," and it traces back to her own background.

Ahuja studied neuroscience and computer science at the University of Toronto, where Geoffrey Hinton's work on neural networks pulled her toward AI. She later worked at Amazon Robotics building robots that build other robots, which is where she met co-founders Will Song and Thomson Yen. Song spent years building sales personalization engines; Yen focused on deep learning approaches to how AI agents interpret human behavior. Together they're betting that a model built around how humans actually perceive and process the world will outperform one built around how humans write about it.

What to Watch

Mirror Particle has closed an angel round and says it's near closing its first venture round, a modest position compared to the hundred-million-dollar raises its competitors have already banked. The company will pitch at Startup Battlefield 200, TechCrunch's competition at Disrupt 2026 in San Francisco from October 13-15, with judging on the afternoon of October 15.

The long-term vision stretches well past brand strategy. Ahuja describes the goal as becoming a "general layer for anticipating human behavior," eventually moving from population-level predictions down to individual-level modeling. Whether a from-scratch foundation model can actually outperform fine-tuned LLMs at scale remains unproven, and Mirror Particle is going up against competitors with vastly deeper war chests. But the underlying argument, that text-trained models are a poor substitute for how humans actually perceive and act in the world, is one worth watching other players in this space wrestle with. As Ahuja puts it: "We just need a better model of humans if we're going to work alongside AI and with each other."