TLDR: The HoPeS (Human-Oriented Perspective Shifting) framework uses large language models (LLMs) to simulate multiple human perspectives in socio-ecological systems. It allows users to role-play as different stakeholders, guided by an AI companion for reflection and insight integration. An experiment showed how a user, acting as a research supplier, faced challenges in influencing policy due to competing stakeholder interests, mirroring real-world researcher-policymaker gaps. The framework aims to enhance understanding of complex systems by enabling narrative-driven and numerical experiments, though challenges like LLM accuracy and role alignment remain.
Understanding how complex systems involving both humans and nature work requires looking at them from many different angles. However, getting these diverse viewpoints from various groups of people can be quite challenging in the real world. To help with this, researchers have developed a new modeling framework called HoPeS, which stands for Human-Oriented Perspective Shifting. This innovative approach uses advanced artificial intelligence, specifically large language models (LLMs), to simulate different human perspectives within these complex systems.
The HoPeS framework is designed to allow users to “step into the shoes” of various stakeholders, experiencing firsthand the differences in how different groups perceive and interact with a system. Imagine being able to see a situation not just from your own point of view, but also from that of a policymaker, a farmer, or an environmental advocate. HoPeS makes this possible through its unique design.
At its core, HoPeS employs LLM-powered agents to represent these different stakeholders. These agents can communicate, reason, and act using natural language, much like humans do. This allows for more sophisticated and realistic simulations of social interactions, including negotiations, persuasion, and even resistance. A key part of the framework is a “simulation protocol” that acts as a guide, helping users move between different perspectives, reflect on their experiences, and ultimately integrate the insights gained from each role.
To demonstrate this concept, a prototype system was developed. This system consists of two main parts: the Perspective-Taking Simulation (PTS) system and the Reflective Learning Companion (RLC) system. The PTS system is where the actual simulation happens. It uses a sophisticated backend model called InsNet-CRAFTY, which combines a network of institutional decision-making agents (powered by LLMs) with a land use model called CRAFTY. This land use model acts as the environment, influencing and being influenced by the agents’ decisions. Within the PTS, there are also special AI assistants (SAA and VAA) that help users analyze data and interact with the system.
The RLC system serves as an AI-powered guide for reflection. After a user plays a role in the simulation, the RLC helps them think deeply about their experiences, the decisions they made, and the challenges they faced. It encourages users to consider what they learned and how they might approach the situation differently. If a user decides to switch roles, the RLC also facilitates this transition, helping them apply knowledge from previous roles to their new perspective. This iterative process of playing roles, reflecting, and transitioning helps users build a more comprehensive understanding of the system dynamics.
In an illustrative experiment, a user with knowledge in land use governance first took on the role of a system observer. In this role, the user simply watched how the LLM agents interacted and how policies influenced land use without intervening. The observer noted that while the agents argued plausibly for their interests, their decisions didn’t always lead to effective outcomes, particularly regarding budget allocation and policy goals. The user felt frustrated by the research supplier’s unclear recommendations in this initial observation phase.
Next, the user assumed the role of the research supplier. In this capacity, the user analyzed data from the land use model and provided technical, politically neutral recommendations for policy adjustments and budget allocation. Despite the user’s efforts to provide clear, data-driven suggestions, the high-level institution (another LLM agent) did not fully implement these recommendations. For instance, a significant discrepancy was observed in budget allocation, where the user recommended a 70% share for agricultural institutions, but the high-level institution implemented only 45%, influenced by opposing arguments from an environmental NGO agent.
This experiment highlighted a common real-world challenge: the gap between researchers and policymakers. Researchers often base their recommendations on scientific evidence, while policymakers must balance conflicting interests and political considerations. The user, acting as the research supplier, experienced frustration and disappointment, realizing that maintaining political neutrality could limit influence. This led to a motivation to explore different strategies for framing policy recommendations to gain more impact, such as emphasizing potential negative consequences if advice isn’t followed.
Also Read:
- Crafting Future Cities: How AI is Redefining Urban Planning
- HAMLET: Crafting Dynamic AI-Driven Theatre Experiences
The HoPeS framework, as demonstrated by this prototype, offers a powerful tool for exploring complex socio-ecological systems. It allows for both narrative-driven and numerical experiments, providing insights into ethical dilemmas, communication subtleties, and culturally embedded values that are difficult to capture with traditional rule-based models. While promising, the development of LLM-driven simulations is still evolving. Challenges remain, including addressing LLM “hallucinations” (generating factually flawed output), ensuring LLMs accurately mimic specific roles, and developing effective ways to integrate diverse perspectives gained from the simulations. Future research will focus on refining the simulation system and protocols, requiring interdisciplinary collaboration to fully realize HoPeS’s potential in fostering a deeper, more holistic understanding of our interconnected world. You can find more details about this research in the full paper available here.
