Complex Social Interaction Benchmarking
Design a Generative Village benchmark in AgentScope to evaluate 10,000+ agents engaging in peer-to-peer trade over dense mesh social networks.
Source paper
Xuchen Pan, Dawei Gao, Yuexiang Xie, Zhewei Wei, Yaliang Li, Bolin Ding, Ji-Rong Wen, Jingren Zhou
Generated by AI. The paper’s authors did not write these.
Design a Generative Village benchmark in AgentScope to evaluate 10,000+ agents engaging in peer-to-peer trade over dense mesh social networks.
Integrate a Cognitive Constraint layer via guided decoding to programmatically restrict language model output vocabulary and reasoning depth for persona fidelity.
Implement a Time Warp optimistic synchronization mechanism in the AgentScope kernel to support cyclic agent dependencies through speculative execution and state rollbacks.
This paper presents AgentScope, a multi-agent platform designed to support very large-scale simulations (up to 1 million agents) by leveraging an actor-based distributed mechanism. The authors introduce tools for automatic workflow conversion, heterogeneous agent background generation, and a web-based management interface, validating the system's scalability through the "guess 2/3 of the average" game.
The core technique is an actor-based distributed computing model (Agha, 1985) that enables agent-level parallel execution and automatic workflow conversion via a `to_dist` function. This is combined with a placeholder mechanism for asynchronous message handling and a configuration pipeline that uses LLMs to generate diverse agent backgrounds based on demographic distributions.