Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster F in Poster Session F: Thursday, August 6, 1:45 – 3:30 pm, Kimmel Center, Shorin & Rosenthal Rooms

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

Botos Csaba¹, Sreejan Kumar², Austin Tudor David Andrews³, Laurence T Hunt³, Christopher Summerfield⁴, Joshua B. Tenenbaum⁵, Rui Ponte Costa³, Marcelo G Mattar⁶, Momchil Tomov⁷; ¹Agitprop AI, ²Columbia University, ³University of Oxford, ⁴UK AI Security Institute, ⁵Massachusetts Institute of Technology, ⁶New York University, ⁷Boston College

Presenter: Sreejan Kumar

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 similar ways? Leveraging a unique dataset of human game learning with concurrent fMRI recordings, we evaluate model-free/model-based reinforcement learning agents, an approximately optimal theory-based reinforcement learning agent, and state-of-the-art large reasoning models (LRMs) on two complementary dimensions: behavioral patterns and predictivity of human brain representations. Using encoding models, we assess how well internal representations from each system predict brain activity in regions previously implicated in theory-based reinforcement learning. We find that LRMs most closely match human behavioral patterns during game learning and predict brain activity in theory-coding regions an order of magnitude better than model-free and model-based agents. Our results establish LRMs as compelling computational accounts of human learning and decision making in complex, dynamic environments.

Topic Area: Methods, Tools, Theory & Neural Coding