ValidAgent – An Open Repository of Validated Generative Agents for Behavioral Simulations and Multi-Agent Systems

AAMAS 2026 Workshop EMAS Submission25 Authors

Keywords: Open Repository, Generative Agents, Large Language Models

TL;DR: We present an open repository for browsing, comparing, and reusing curated, empirically validated generative agents in behavioral simulations and multi-agent systems.

Abstract: Large Language Models are increasingly used to create generative agents for behavioral simulations and multi-agent systems, as they allow researchers to build entities that exhibit seemingly human-like behavior with relatively low effort. Although generative agents are frequently parameterized using empirical data, systematic empirical validation remains scarce: there is no established practice of benchmarking agent behavior against human reference data, nor a common standard for reporting validation in a way that allows comparison across studies. No dedicated infrastructure exists to support the collection, documentation, and reuse of validated agents. As a result, generative agents are difficult to transfer across studies, hard to reproduce, and lack comparable validation – ultimately limiting cumulative scientific progress. In response, we present ValidAgent, a prototype of an open, web-based repository that provides a common interface for researchers to browse, compare, and select from a curated collection of generative agent profiles. The platform emphasizes empirical validation, transparency, and reproducibility by requiring each agent profile to include structured documentation covering its design rationale, behavioral traits, underlying assumptions, intended scope of application, and validation results against human reference data. To demonstrate the platform’s utility, we provide an initial set of empirically validated agents grounded in the die-roll honesty paradigm, a widely used experimental setup for studying dishonest behavior. Building on this, we aim to foster a researcher-driven ecosystem in which generative agents can be contributed, peer-reviewed, and benchmarked – ultimately establishing citable generative agent sets. In doing so, this initiative contributes to a more transparent and cumulative multi-agent simulation practice.

Paper Type: Tools / Testbeds / Demo paper

Demo: Yes, we would love to present a demo.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 25