Modeling Human Complexity

How Large Behavior Models and Social Simulators Are Rewiring AI

Artificial intelligence is evolving - moving beyond the mechanics of language processing to take on the richness of human behavior and social dynamics. This expansion is giving rise to Large Behavior Models (LBMs): systems designed to interpret, predict, and simulate the intricate patterns of human interaction.

An example of a compelling recent developments in this space is YuLan-OneSim, a sophisticated social simulator that places LLM-driven agents into dynamic, multi-agent environments. Far from being a novelty, YuLan-OneSim is an ambitious leap toward modeling the subtleties of collective human behavior at scale.

A Sandbox for Social Simulation

YuLan-OneSim offers researchers, policymakers, and designers a robust framework for exploring how behaviors emerge, evolve, and influence social systems. It’s a lab for behavioral experimentation - without the constraints of physical time, space, or ethics.

Key features include:

• Code-Free Scenario Creation

  Users input natural language descriptions to define simulation parameters, and the platform automatically generates the code needed to run those scenarios. This opens the door for researchers from sociology, economics, and psychology who may not have formal programming backgrounds.

• 50 Pre-Built Scenarios Across Eight Domains

  These domains include Economics, Psychology, Politics, Sociology, Law, Health, Communication, and Organizational Behavior. The default scenarios allow for exploration of real-world issues - from opinion polarization to epidemic spread - within realistic digital populations.

• Evolvable Simulations

  The system incorporates feedback (from users or internal agents) to continuously refine its models, enhancing both realism and accuracy. This allows for a kind of “social learning loop” where simulations grow more sophisticated over time.

• Massive Scale, Reliable Results

  YuLan-OneSim supports up to 100,000 simulated agents, enabling experimentation with entire social networks, cities, or systems.

• An AI Social Researcher

  With agent collaborators for scenario design and technical reporting, YuLan-OneSim can automate key parts of the research pipeline - analyzing data, generating reports, and iterating on findings using scientific protocols like ODD (Overview, Design Concepts, Details).

The system has already shown promise in areas such as validating Axelrod’s cultural diffusion model or modeling the housing price distribution in Brazil. It gives researchers a lens into emergent human behavior - one that can be adjusted, scaled, and interrogated.

Building Behavioral Intelligence for AI

As technologies like YuLan-OneSim push the boundaries of what AI can simulate, the demand for realistic, behaviorally grounded data becomes increasingly urgent.

Concrete is developing a Large Behavior Model (LBM) - an AI system designed to embed a deep understanding of human motivation, decision-making, and social context directly into intelligent systems. This LBM will be rooted in nearly two decades of behavioral research and real-world insights gleaned from human-centered research and design, communications, and digital product work.

Rather than relying on transactional data or surface-level engagement metrics, Concrete’s LBM will draw from a foundation of proprietary research into human relationships, social communication, and emotional drivers as well as decades of aggregated learnings and generalized knowledge from nearly twenty years of human-centered research and design experience. This will enable AI systems not just to recognize patterns, but to understand why they form - adding the dimension of intention, bias, and context to machine decision-making.

The vision is to create systems that anticipate behavior, simulate meaningful interaction, and respond not just with intelligence - but with emotional and social fluency.

Concrete’s future model aims to go beyond conventional AI by incorporating psychological nuance, cognitive heuristics, and environmental context. The vision is to create systems that anticipate behavior, simulate meaningful interaction, and respond not just with intelligence - but with emotional and social fluency.

As simulations like YuLan-OneSim continue to evolve, they will need this kind of nuanced, behaviorally rich data to stay grounded in real-world human complexity. Concrete’s aim is to become a key supplier of that intelligence - a future node in what could be described as the behavioral supply chain of advanced AI systems.

Simulation Meets Behavioral Signal

Simulations are only as accurate as the assumptions and data they are built upon. The power of YuLan-OneSim lies not just in its technical capabilities, but in its potential to become a proving ground for behavioral hypotheses - testing how collective dynamics emerge from individual decisions, and how those behaviors can be influenced, shaped, or rebalanced.

Concrete envisions becoming an enabler for this kind of high-fidelity simulation. By supplying behavioral data grounded in lived experience - rather than abstract metrics - they plan to help simulation platforms like YuLan-OneSim better reflect reality.

Concrete’s behavioral intelligence will not just be about improving AI - it will be about making AI socially literate. This means creating systems that can anticipate friction, model trust, and understand the sometimes illogical ways people act when under pressure, in groups, or in unfamiliar contexts.

Toward AI That Understands Us

The convergence of simulation and behavior is more than a technical milestone - it’s a philosophical one. We are beginning to build systems not only for humans, but that understand what it means to be human. That’s a different kind of intelligence - one that’s contextual, emotional, and deeply rooted in shared experience.

YuLan-OneSim offers a glimpse into what’s possible when advanced simulation platforms model the complexities of human behavior. Concrete, meanwhile, is building the kind of deep behavioral data infrastructure that models like these will increasingly depend on.

As this field advances, the ability to simulate behavior at scale and depth will redefine everything from product design to policy forecasting.

They represent a new horizon for AI - one where predictive models aren’t just functional, but fluent in the messy, beautiful logic of human life. As this field advances, the ability to simulate behavior at scale and depth will redefine everything from product design to policy forecasting. But simulation alone isn’t enough. We need the human thread woven in. We need models built on insight, not just input.

Concrete is building for that future. One where behavioral data powers machines that don’t just process our actions - but understand them.

Read the full submission on YuLan-OneSim.