Frontier LRMs Match Human Game Learning and Brain Activity
A recent research paper available on arXiv (2605.08019) evaluates cutting-edge Large Reasoning Models (LRMs) alongside deep reinforcement learning agents and a Bayesian theory-driven agent in intricate video game learning scenarios. By analyzing a dataset of human gameplay paired with simultaneous fMRI scans, the researchers discovered that LRMs align more closely with human behavior during game exploration. Additionally, they demonstrated that LRMs can predict brain activity significantly more accurately than reinforcement learning methods across various cortical areas.
Key facts
- Study uses dataset of complex human gameplay with concurrent fMRI recordings
- Participants learned novel video games requiring rule discovery, hypothesis revision, and multi-step planning
- Compared frontier LRMs against model-free and model-based deep RL agents and a Bayesian theory-based agent
- Frontier LRMs most closely match human behavioral patterns during game discovery
- LRMs predict brain activity an order of magnitude better than RL alternatives
- Study published on arXiv with ID 2605.08019
- Research focuses on abstract knowledge learning and flexible deployment in novel environments
- Joint evaluation of models on gameplay, human learning behavior matching, and brain activity prediction
Entities
Institutions
- arXiv