Ententsociety is a cutting-edge open source framework designed to simulate a large number of agents, each powered by a large language model (LLMS) to realistically simulate the complex interactions found in human society. Leveraging powerful distributed processing techniques (especially Ray), the project implements simulations involving tens of thousands of simultaneous active agents, each embedded in detailed, realistic environments that capture social, economic and mobility behaviors.
Key Features
Large scale and fast performance
- Supports a large number of people: The framework shows simulations of up to 30,000 agents, exceeding the wall lock time, that is, on virtual society, it runs faster than real-time time 1.
- Parallelize with rays: EverentSociety uses a ray framework to manage large-scale parallel execution of agents, which is critical to handling large-scale and nondeterministic interactions.
- Effective resource usage: By grouping agents and sharing network clients in groups, the framework greatly reduces memory and connection overhead, overcoming port and memory bottlenecks common in scaling distributed simulations.
The real social environment
Strengthening differentiating oneself by integrating highly realistic feedback and constraints enables agents to act in a way that reflects the actual social system.
- Urban Space: Combined models of real-world map data (e.g. from OpenStreetMap), road network, points of interest, and mobility (walking, driving, public transportation), updating the second of each simulation.
- Social Space: Agents form evolving social networks and engage in online and offline social interactions. Messaging (including content auditing and user blocking) is modeled to simulate communication patterns in social media and real-world.
- Economic Space: Implement employment, consumption, banking, government (tax) and macroeconomic reporting, which are all driven by agency decisions. Agents must balance income and expenditures and simulate realistic economic behavior.
Construction and Technology
Parallel interaction engine
- Group-based distributed execution: Agents are divided into groups managed by Ray “Actors” that optimize resource usage by using asynchronous network requests for connection reuse while maintaining high parallelism.
- High-performance messaging: Leveraging Redis’s bar/sub-function, agents communicate effectively and support agent agents and user agents (external program) interaction.
- Time alignment mechanism: The framework synchronizes proxy and environment progress, ensuring consistent and repeatable simulations despite changes in processing time of LLM API calls.
- Comprehensive Utilities: Simulate records (by PostgreSQL and local file storage), metric records (MLFLOW), and GUI for experimental creation/management and result visualization.
Quantitative results
Scalability and speed
- Faster than real-time: In a deployment using 24 NVIDIA A800 GPUs, simulations of 30,000 agents achieve faster operation speeds (e.g., all execution speeds are faster than the equivalent real-world elapsed time).
- Linear Scaling: Performance is linearly scaled through computing resources; increasing the LLM service GPU can achieve higher simulation throughput until service limits of the backend language model.
- Sample metrics: In the largest experiment (30,000 agents, 8 groups), the average agent was completed in 252 seconds, kept in real time, and completed with 100% LLM call success rate. The environmental simulation and message delivery time is still much lower than the LLM inference time, thus confirming the computing efficiency of the system.
The impact of the real environment
- The authenticity of agent behavior: Compared to LLM-Prompt-based “text simulators” and various generative trajectory benchmarks, a combined reality-based environment simulator can significantly improve the authenticity and human style of proxy behavior.
- Experience benchmark: On measures such as gyro radius, location of daily access, and distribution of behavioral intentions, LLM agents with environmental support greatly exceed the timely and classical model baselines, closely matching real-world data.
Use cases and applications
An open design and configurable environment makes the additive society a powerful tool:
- Social Science Research: Research on social models, emerging phenomena, mobility and information dissemination.
- Urban Planning and Policy Analysis: Before realistic deployment, evaluate interventions in simulated environments.
- Management Science: Model organizational dynamics, labor changes and economic behavior.
in conclusion
Reinforcement is the first open source framework to simulate social interactions effectively and realistically at an unprecedented scale. It integrates LLM-driven agents with parallel data-driven environments into key tools for computing research and practical decision support to understand complex social dynamics.
Check Paper and project. All credits for this study are to the researchers on the project. Also, please stay tuned for us twitter And don’t forget to join us 100K+ ml reddit And subscribe Our newsletter.
Asif Razzaq is CEO of Marktechpost Media Inc. As a visionary entrepreneur and engineer, ASIF is committed to harnessing the potential of artificial intelligence to achieve social benefits. His recent effort is to launch Marktechpost, an artificial intelligence media platform that has an in-depth coverage of machine learning and deep learning news that can sound both technically, both through technical voices and be understood by a wide audience. The platform has over 2 million views per month, demonstrating its popularity among its audience.
