What Do Agents Think One Another Want? Level-2 Inverse Games for Inferring Agents' Estimates of Others' Objectives

A team from UT Austin and Stanford has published a novel research paper introducing 'Level-2 Inverse Games,' a framework designed to model how AI agents reason about each other's goals in strategic interactions. The work, led by Hamzah I. Khan, Jingqi Li, and David Fridovich-Keil, tackles a fundamental flaw in existing 'inverse game theory' models. Current 'Level-1' approaches assume all agents in a system share a complete and accurate understanding of each other's objectives. The authors prove this assumption breaks down in real-world, decentralized scenarios like autonomous driving or bargaining, where agents act based on incomplete or conflicting views of what others want.

The new 'Level-2' framework explicitly asks and answers the question: 'What does each agent *believe* about other agents' objectives?' The researchers demonstrated that failing to model these second-order beliefs leads to significant prediction errors, which they characterized theoretically using data from linear-quadratic games. They developed an efficient gradient-based algorithm to solve this inherently non-convex inference problem. In experiments on a synthetic urban driving scenario, their approach successfully uncovered nuanced strategic misalignments that traditional Level-1 methods completely missed, providing a more accurate model of complex multi-agent behavior.