The Work
Csabi frames teaching as a theory-of-mind problem with three pillars: estimate what the learner knows, model how they learn, and find the Goldilocks zone. Reason to Play is pillar one. 32 people learned ARC-AGI like games with no instructions in an fMRI scanner. Eight frontier reasoning models played the same games with no prior training.
What Worked
Reasoning models predicted held-out human brain activity an order of magnitude better than RL and Bayesian baselines across cortical and subcortical regions, and matched human discovery speed where deep RL needed 100 to 1,000x more interaction. Same-architecture untrained controls sat near chance (6.4 to 6.9x gap, p below 10⁻¹¹).
What Has Been Achieved
The models discovered the rules at speeds similar to humans, with 100 to 1,000x less interaction than SOTA agents, and their hidden states predicted brain activity an order of magnitude better. The paper drew orals at ICLR, CogSci and CCN, 35+ invited talks and a Digital Brain Project award to collect data for the continuation of Csabi's research program.
What Was Learned
Training agents from scratch to match brain data never leaves the random-initialisation floor, a dead end for the cohort. And the alignment tracks the model’s representation of the game state rather than its planning, so the neural signature of active learning itself is still uncaptured.
About
Csabi holds a PhD in core ML (focusing on Continual and Self Supervised Learning) from University of Oxford with collaborations at Apple, Intel and Meta. Prior to his fellowship he worked on Multi-Modal World Models as an ML engineer for Silent Creek, a High Frequency Trading firm.
Lab & Advisors
Chris Summerfield
University of Oxford
Professor of Cognitive Neuroscience at the University of Oxford, Research Director at the UK AI Security Institute
Chris' work focuses understanding the cognitive and neural mechanisms that underlie human learning and decision-making, and on studying the impacts of AI on society. His research bridges the fields of cognitive science, neuroscience, and artificial intelligence. He is particularly interested in how insights from human cognition can inform the development of more advanced and safer AI systems. Chris also lead the Human Information Processing (HIP) lab in the Department of Experimental Psychology at the University of Oxford.
Rui Ponte Costa
Group Leader, Centre for Neural Circuits and Behaviour, Department of Physiology, Anatomy and Genetics
Rui Ponte Costa leads the Neural & Machine Learning Group at Oxford, bringing together neuroscience, psychology and machine learning.
