A small behavior model that beats frontier LLMs at predicting human reactions

Language is what people say. Behavior is what they choose. We’re building AI for the second.

Read the research

A new lens for artificial intelligenceLanguage. Worlds. Now, human behavior.

AI can write the message. It can model the environment. Our question is different: what will the person do? Cairo Rosetta is a specialized behavior model built to predict individual choices—not just produce convincing answers.

LLMsThe language layer.

Large language models generate and interpret language. Their general-purpose capabilities make them useful across many tasks.

World modelsThe world layer.

World models represent environments and how they change. Their focus is the dynamics of a world and its possible future states.

LBMs · Cairo RosettaThe behavior layer.

Our focus is the person inside that world. Given a profile, a situation and possible actions, which choice is that individual likely to make?

Our mission: build the world’s first Large Behavior Model. This is an ambition, not a verified claim of historical priority. LLMs, world models and behavior models are complementary lenses, not a ranking of intelligence.

A population average can hide the very differences you need to understand. Rosetta conditions its prediction on who the person is, what is at stake and what they can actually choose.

Psychometric and cognitive traits describe the individual—not a generic role-play prompt.

02 / The situationContext + possible actions

The scenario, its constraints and a defined set of alternatives frame the decision.

03 / The predictionA structured choice

Rosetta selects the action it estimates best matches that person in that situation.

Behavioral simulation becomes practical when the decision engine is fast and economical enough to repeat. Rosetta is designed for low-latency, low-marginal-cost predictions across many profiles and scenarios.

Our preliminary research reports lower mean latency than every frontier language model evaluated. Dedicated behavioral inference keeps the focus on the choice.

Rosetta led the evaluated models in point accuracy. The direction is encouraging; a statistically conclusive advantage has not yet been established.

Dedicated inference avoids a paid external LLM API call for each Rosetta prediction. Infrastructure still costs money; this is not a claim of free or zero-cost simulation.

These comparisons describe preliminary Cairo research, not a guarantee for every workload. End-to-end speed, accuracy and cost depend on the task and the surrounding simulation pipeline.

The horizonUnderstand the choice before you make your move.

Our long-term ambition is a behavior layer for decisions about products, communication and social systems. These are directions for development and validation—not claims that every domain is already solved.

Explore how different people might respond to an offer, navigate an experience or weigh alternatives.

Investigate the gap between what a message says and the decision it may inspire.

Study how profiles and context interact when people face trade-offs, uncertainty and negotiation.

Cairo Rosetta is a proprietary, specialized behavior model developed by Cairo Labs. It uses an individual’s psychometric profile, the context of a situation and a catalog of available actions to estimate which action that person is most likely to take. Cairo calls this direction a Large Behavior Model, or LBM.

An LBM is our framing for AI focused on modeling human decisions, rather than primarily generating language or modeling environmental dynamics. These categories overlap; they are not mutually exclusive or a universally standardized taxonomy. Cairo’s mission is to help establish the behavior layer as a category in its own right.

Rosetta achieved a higher point accuracy than the frontier language models evaluated in our preliminary research. The evidence is encouraging, but the accuracy advantage is not statistically conclusive. It does not establish universal superiority across tasks, audiences or model versions.

No. Rosetta avoided external language-model API charges in the reported research, but infrastructure and other simulation stages still have costs. The paper does not establish an all-in inference price or a commercial Rosetta API price. Cairo workspace simulations use the platform’s published credit plans.

No. Behavioral predictions are estimates, not observed decisions or guaranteed outcomes. Use simulated signals to explore hypotheses and prioritize what to test with real people. Broader validation across domains remains part of our research agenda.

Explore Cairo’s audience simulation workspace. Turn your next question into something you can test.