LRMs align with human brain and behavior 10x better than deep RL

Researchers from several leading institutions—including DeepMind, Princeton, Oxford, and NYU—published a paper titled “Reason to Play” that directly compares frontier Large Reasoning Models (LRMs) with deep reinforcement learning (RL) agents on how well they mimic human learning and brain activity. They used a dataset of complex video game play from human subjects whose brains were scanned with fMRI. The LRMs were evaluated on three criteria: ability to play the games, match human learning behavior, and predict neural signals. Against model-free and model-based RL agents, the LRMs came far closer to human behavioral patterns during the discovery phase of gameplay and predicted brain activity an order of magnitude better—a result robust to permutation controls.

The most striking finding is that the brain alignment is driven by the LRM’s in-context representation of the game state itself, not by its downstream planning or reasoning. This suggests that these models encode information in a way that aligns with how human brains represent task structure. By establishing LRMs as a strong computational account of human learning and decision making in naturalistic environments, the work opens the door to using them as proxies for human cognition—with implications for neuroscience, AI safety, and human-AI interaction design.