Self-calibrating generative agents

Summarise with VIRAT
Authors: Ting Huang
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This project introduces a self-calibrating LLM-agent evacuation framework that integrates a multi-agent in a grid-based emergency simulation with an online prompt calibration loop to achieve an adaptable and explainable decision-making process under dynamic crowd conditions. The system models exit-choice behaviour and congestion dynamics while calibrating agent behaviour towards prescribed distribution. At the agent reasoning layer, each simulated agent follows a four-stage interaction process. First, the agent assesses potential threat from its environment; second, it evaluates exits choices based on distance and crowding; third, it selects an exit informed by neighbouring agents; and finally, it moves under simulated physical constraints such as obstacles. This design turns environment signals into interpretable decisions, making agent behaviours explainable. At the calibration layer, an online self-calibrating prompt loop continuously monitors the distribution of agents across the grid and compares it against predefined target distributions at each step. When deviation exceeds a certain threshold, the system injects targeted guidance into the agent’s decision process to encourage choosing the desired exit. Together, these components form a unified pipeline linking environment dynamics, agent reasoning, and feedback control loop to support a transparent and explainable evacuation simulation.

Fields of Study & Subjects

Evacuation and Crowd Dynamics Simulation Techniques and Applications Multi-Agent Systems and Negotiation

Publication details

Journal / source
DR-NTU (Nanyang Technological University)
Publication date
1 January 2026
Language
EN