Frontier Large Reasoning Models (LRMs) demonstrate human-like learning patterns in complex game scenarios and predict brain activity an order of magnitude better than reinforcement learning approaches, according to new research combining fMRI recordings with AI performance analysis. The study establishes LRMs as compelling computational models of human cognition by incorporating neuroscientific validation rather than relying solely on behavioral metrics.
Superior Performance Across Behavioral and Neural Metrics
Researchers used complex video game scenarios paired with fMRI brain recordings to compare how different AI systems and humans approach novel learning environments requiring rule discovery, hypothesis revision, and multi-step planning. The study compared frontier LRMs against model-free and model-based deep reinforcement learning agents, plus a Bayesian theory-based agent.
The findings revealed two critical advantages for LRMs:
- Behavioral Alignment: LRMs most closely replicated human behavioral patterns during game discovery and hypothesis revision across the novel gaming tasks
- Brain Activity Prediction: These models predicted brain activity an order of magnitude better than both reinforcement learning alternatives across cortical and subcortical regions
Game State Representation Drives Brain Alignment
Targeted manipulations revealed that brain alignment stems from the model's representation of game state rather than its downstream reasoning mechanisms. This mechanistic insight provides a deeper understanding of how these models align with human cognition at a fundamental level, distinguishing between state representation and planning processes.
Bridging AI Capabilities and Cognitive Science
The research is notable for incorporating neuroscientific data to validate AI alignment with human reasoning. Rather than assuming behavioral similarity indicates cognitive similarity, the authors used brain recordings to confirm that LRMs capture human-like reasoning in complex, naturalistic environments.
The paper, "Reason to Play: Behavioral and Brain Alignment Between Frontier LRMs and Human Game Learners," was published on arXiv on May 8, 2026, by Botos Csaba and colleagues. The work suggests that frontier reasoning models have advanced beyond pattern matching to exhibit cognitive processes that mirror human learning and planning.
Key Takeaways
- Frontier Large Reasoning Models predict human brain activity an order of magnitude better than reinforcement learning approaches in game learning tasks
- LRMs demonstrate superior behavioral alignment with human patterns during rule discovery and hypothesis revision
- Brain alignment stems from the model's game state representation rather than downstream reasoning processes
- The research validates LRMs as computational models of human cognition using neuroscientific data, not just behavioral metrics
- The study bridges AI capabilities with cognitive science by testing models in complex, naturalistic learning environments