Modeling Others' Minds as Code
9 hours ago
- #Program Synthesis
- #Human Behavior Modeling
- #Artificial Intelligence
- Accurate prediction of human behavior is crucial for safe human-AI collaboration.
- Existing approaches for modeling people are often data-hungry and brittle.
- Everyday social interactions may follow predictable patterns or 'scripts' to minimize cognitive load.
- Proposes modeling routines as behavioral programs in code rather than policies based on beliefs and desires.
- Introduces ROTE, an algorithm combining LLMs for synthesizing behavioral programs and probabilistic inference for uncertainty.
- ROTE tested in gridworld tasks and a household simulator, outperforming baselines by up to 50% in accuracy and generalization.
- Treats action understanding as a program synthesis problem for efficient human behavior prediction.