From Interaction to Abstraction: Using Behavior and Brain Activity to Evaluate How AI Systems Learn Games

Botos Csaba, Sreejan Kumar, Austin Tudor David Andrews, Rui Ponte Costa, Marcelo G Mattar, Momchil Tomov

Keywords: reinforcement learning, large reasoning models, video games, human abstraction, bayesian models

TL;DR: We evaluate how Large Reasoning Models play complex videogames that require planning by comparing to human behavior and brain activity,

Abstract: Humans rapidly learn abstract knowledge when encountering novel environments and flexibly deploy this knowledge to guide efficient and intelligent action. Can modern AI systems learn and plan in a similar way? Leveraging a unique dataset of human gameplay with concurrent fMRI recordings, we compare model-free and model-based reinforcement learning agents, bayes-optimal model-based agents, and frontier Large Reasoning Model. We evaluate models on task performance and behavioral alignment, then use brain activity as a benchmark to probe the internal representations these models construct. Specifically, using encoding models, we assess how well each system’s internal representations predict brain activity in regions previously implicated in theory-based reasoning. We find that the Large Reasoning Model most closely matches human behavioral patterns during game discovery and predicts brain activity in theory-coding regions an order of magnitude better than both model-free and model-based alternatives. Our results shed light on the computational principles underlying human-like rapid learning and planning.

Paper Type: New Full Paper

Submission Number: 46