Building Simulation Environments for Computational Organizational Design

ICLR 2026 Conference Submission13770 Authors

Keywords: Reinforcement Learning environment, LLM agent, simulation, Organizational Science, Adaptive Management Policies, Multi-Objective Optimization, Multi-Agent Systems, Clinical Trials

TL;DR: This paper introduces the Organizational Design Problem with a blueprint to build environments that enable the development of management policies. We showcase this by releasing a Clinical Trial Environment, which models a drug development program.

Abstract:

Organizational success depends less on individual brilliance than on how teams are structured, coordinated, and adapted. Yet organizational design remains a grand challenge in computational science, and machine learning lacks tools to address it. We introduce the Organizational Design Problem (ODP): learning a management policy that configures team composition, communication, and autonomy to achieve multi-objective goals under structural constraints.

A main obstacle to developing machine learning for the ODP is the lack of suitable Organizational Simulation Environments (OSEs) in which such policies can be learned and evaluated. While organizational design is a general task as organizations are a universal feature of social and economic life, each organization is unique in its purpose, internal constraints, and external surroundings. Acknowledging this specificity, we propose an OSE blueprint: it defines the core components shared by all organizations while allowing adaptation to diverse contexts. In this framework, fixed LLM agents simulate realistic human roles and communicate via natural language within a mechanistic, temporally grounded simulation.

Applying this blueprint, we present the Clinical Trial OSE, which captures the high-stakes, multi-stakeholder process of drug development. Using this environment to benchmark pre-trained LLMs, we show that they can guide organizations to successfully complete trial programs. Although current models remain less efficient than humans, our study opens the path toward specialized models that could one day outperform humans in systematically solving the Organizational Design Problem.

Primary Area: datasets and benchmarks

Submission Number: 13770