A project directed by Max Kleiman-Weiner at the University of Washington aims to  understand how and why  human caregiving works in computational terms. The capacity to care for others represents one of humanity's most profound and distinctive forms of intelligence, yet no formal models capture the sophisticated reasoning underlying caregiving decisions. By reverse-engineering caregiving principles, the project seeks to illuminate how different forms of intelligence solve the complex optimization problem of fostering autonomy in developing agents.

The hypothesis frames caregiving as a computational problem of helping learners with latent capabilities achieve independence and autonomy. Autonomy is defined as the learner's capacity to achieve goals across diverse future environments, involving three components: capability, independence, and adaptability.

The team will model caregiving as a multi-stage optimization problem where caregiving actions during development maximize the learner's utility during a later autonomous period when the caregiver is absent. This framework differs from mere helping by optimizing for future outcomes rather than immediate satisfaction. It incorporates cultural beliefs about development and future environments and focuses explicitly on autonomy enhancement.

The project has three main aims:

• Computational Framework: Develop and simulate the autonomy-centered model, comparing it to alternatives (e.g., empowerment, teaching) in grid-world environments, addressing scalability via approximations such as curriculum learning and safe exploration.
• Empirical Validation: Test predictions through online experiments with caregiver decisions in 2D hazard/reward environments; analyze parent-child “takeover” behaviors (intervening vs. allowing struggle); and study naturalistic data from child-perspective headcam recordings (BabyView dataset).
• AI Alignment: Apply caregiving principles to AI systems (e.g., tutoring agents) designed to enhance human autonomy rather than create dependency, measured by user performance after AI assistance ends.

If successful, the project will produce the first formal computational theory of human caregiving, validated through cross-cultural empirical studies and implemented in prototype AI systems that enhance, rather than replace, human capabilities.