Reason to Play
Behavioral and Brain Alignment Between Frontier LRMs and Human Game Learners
Botos Csaba1*,
Sreejan Kumar2,3*,
Austin Tudor David Andrews1,
Laurence Hunt1,
Chris Summerfield1,
Joshua B. Tenenbaum4
Rui Ponte Costa1+,
Marcelo G. Mattar3+,
Momchil Tomov5+
1University of Oxford, 2Columbia University, 3New York University, 4MIT, 5Harvard University — * Equal first authors, + Equal senior authors
Summary
Humans are remarkably good at picking up new environments: drop someone into an unfamiliar video game and within minutes they are forming hypotheses, testing them, and revising their understanding of how the world works. Can modern AI do the same? We take a set of simple grid-world video games, written in a compact language called VGDL, and have both humans (32 participants, scanned with fMRI) and a suite of frontier Large Reasoning Models play them from scratch, with no rules provided. We then ask two questions: Do the models learn the way humans do? And do they build similar internal representations?On the behavioral side, the best LRMs match human learning curves remarkably well: they discover game rules in roughly the same number of steps humans need and progress through difficulty levels at a comparable rate. On the neural side, the hidden states of these models predict human fMRI BOLD responses across cortical and subcortical regions, the first demonstration that LRM representations systematically align with human brain activity during active game learning. Through targeted ablations we show that what drives the brain alignment is the model's in-context representation of the sequence of game states.
Do Models Represent the Games as Humans do?
The games on this page are the exact same environments used in the experiment, including the randomized colour assignment. Nobody, neither the human participants nor the LLMs, was told what the rules are. On the left side you see a human participant's gameplay (top left), their neural representation evolving during in-context learning (bottom left), while on the (top right) you see what the LLM sees and in (bottom right) how their representation formed for the same gameplay.At the top you see a human participant's gameplay, in the 2nd panel what the LLM sees, in the 3rd panel their neural representation evolving during in-context learning, and at the bottom how the model's representation formed for the same gameplay.How does the model think while it plays?
Working with language models has one big advantage over silent neural networks: we can read the model's reasoning. In the copied-reasoning condition, the model's hidden chain of thought is exposed at each step. Below is one such run: DeepSeek-V3.2 playing Bait level 3 with no rules provided. Step through the gameplay yourself, then read the curated reasoning trace underneath. The model proposes candidate object identities, tests them against observed state changes, and revises its understanding when predictions fail, in a way that reads strikingly like how human participants describe their own learning when thinking aloud. Click on the game cards to start the replay, or the titles to watch the full traceAbove is the model's actual gameplay: each frame is one action chosen from the same observation a human participant would see, and the text panel prints the complete chain of thought the model wrote down before picking it. For illustrating if there are any similarities between human Response Times (RT), we stagger the gameplay here by waiting until the model Reasoning Traces (RT, convenient huh?) is completed in the text panel. Step through the run, open the Game Description flap to peek at the rules the model never saw, and watch how its hypotheses about object identities evolve across attempts.
The figure below distils the same run into a short, readable narrative: only the moments where the model commits to or revises a hypothesis are kept, and arrows show how each observation feeds back into the next decision. Read it as a "think aloud" transcript: the live viewer is the primary source, the figure is the curated commentary.
Reasoning trace (Bait, level 3).
Curated excerpts from the DeepSeek-V3.2 run shown above.
The model proposes candidate object identities, verifies them against observed state changes, and adapts its strategy when predictions fail.
How well do the matches hold up?
The viewer above and the reasoning trace are striking qualitatively, but a research claim needs more than a single visually compelling case. We need to quantify two things: how closely the models behave like human game-learners, and how closely their internal representations match human brain activity.Quantifying behaviour
To compare learning behaviour, we measure two things: (i) rule-discovery time, the number of steps each agent takes to first satisfy the win condition of a game; and (ii) curriculum progression, how far through nine difficulty levels each agent advances under blocked advancement (two consecutive wins required to move on). Human participants, deep-RL baselines (DDQN, EfficientZero, EMPA) and eight frontier LRMs all play the same games under the same conditions.
Learning efficiency and capability.
Top: discovery-time distributions (KDE); LRMs overlap with humans while deep-RL baselines are shifted far right.
Bottom: curriculum progression under blocked advancement (2 consecutive wins required to advance); the best LRMs track the human staircase closely.
Quantifying representational similarity
To compare internal representations, we extract hidden-state activations from each LRM during gameplay and use them as regressors in a voxelwise encoding model that predicts the human fMRI BOLD signal, with separate regularisation for the model features and the nuisance regressors (game/level identity, button presses, time). We then average best-layer Pearson correlations within functional region groups. Targeted ablations (prompt-only, shuffled-features, random-init controls) confirm that the signal comes from the model's in-context representation of the game state, not from surface-level prompt statistics or chance correlations.
Brain encoding accuracy by region group.
Best-layer Pearson correlation averaged across voxels within each functional group.
LRM features (right) consistently outperform deep-RL baselines (left) across cortical regions.
Where can this go?
Two threads suggest themselves naturally from the results above.Closing the active-learning gap
The brain alignment we report is mostly about what the world is: the structure of game states, encoded in the model's hidden activations. It is much harder to read off the brain signal of active learning and planning itself — the moments where a participant forms a hypothesis, decides to test it, and revises when the prediction fails. Language representations are surprisingly good proxies for perceptual content (Marjieh et al. 2022; Huth et al. 2016), but cognitive operations like search, planning and retrieval leave a different kind of footprint (Fedorenko et al. 2024). A natural next step is to commit to a richer harness: external memory, explicit reasoning scratchpads, cognitive-architecture choices that mirror what humans do (Sumers et al. 2023), and a way to align the model's own gameplay trajectory back to the individual participant's. Once models can actively engage with the environment in a way that matches a specific human run, we expect the encoding analysis to recover signal it currently misses.A standardised testbed for theory comparison
Once the active-learning gap is closed, this paradigm offers cognitive science something it has rarely had: a way to test competing accounts of human learning and planning jointly against both behaviour and the brain, on the same humans, the same games, the same TR axis. Candidate accounts — Bayesian program induction, schema-based reasoning, language-as-thought, planning-by-search — can each be operationalised as a harness on a frontier LRM and scored simultaneously on rule-discovery time, curriculum progression, and per-voxel encoding accuracy. The interactive surfaces above (the replay viewer in Section B, the representation viewer in Section A) are a first attempt to make those comparisons inspectable, not buried behind a regression coefficient table.Caveats
Two things bound the scope of our claims. First, estimating per-voxel noise ceilings is genuinely hard when every gameplay trajectory is unique, so cross-region comparisons of absolute encoding accuracy mostly reflect regional signal-to-noise; the cleanly interpretable claim is the relative ordering across model classes, not the absolute numbers. Second, LRMs retain general priors from their (opaque) training corpora that we cannot fully control for. Every model class in our comparison has access to some game-relevant information through one channel or another — deep-RL receives weight updates from specific levels, EMPA has the ground-truth VGDL ontology, LRMs have their pretraining mix — but the channels are not equivalent. We see this as motivating, not deflating: the profile of LRM priors is closer to how a human approaches a novel game than from-scratch training, and that's exactly why the comparison is informative in the first place.Citation
@inproceedings{botos2026reasontoplay,
title = {Reason to Play: Behavioral and Brain Alignment Between Frontier {LRMs} and Human Game Learners},
author = {Botos, Csaba and Kumar, Sreejan and Andrews, Austin Tudor David
and Hunt, Laurence and Summerfield, Chris and Tenenbaum, Joshua B.
and Ponte Costa, Rui and Mattar, Marcelo G. and Tomov, Momchil},
booktitle = {Advances in Neural Information Processing Systems},
year = {2026},
note = {Accepted at NeurIPS 2026},
url = {https://openreview.net/forum?id=Y1oX1yuaWM}
}