🧬 biology

Reason to Play: Behavioral and Brain Alignment Between Frontier LRMs and Human Game Learners

This paper demonstrates that frontier Large Reasoning Models (LRMs) significantly outperform traditional reinforcement learning agents in both matching human gameplay behavior and predicting brain activity during novel game learning, establishing them as superior computational models of human learning and decision-making in complex environments.

Original authors: Botos Csaba, Sreejan Kumar, Austin Tudor David Andrews, Laurence Hunt, Chris Summerfield, Joshua B. Tenenbaum, Rui Ponte Costa, Marcelo G. Mattar, Momchil Tomov

Published 2026-05-11

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Botos Csaba, Sreejan Kumar, Austin Tudor David Andrews, Laurence Hunt, Chris Summerfield, Joshua B. Tenenbaum, Rui Ponte Costa, Marcelo G. Mattar, Momchil Tomov

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are teaching a group of very different students how to play a brand-new, complex board game. You have three types of students:

The "Brute Force" Learners (Deep RL): These students try every possible move, fail thousands of times, and slowly memorize the rules through sheer repetition. They are like a dog learning a trick by getting a treat every time it sits correctly.
The "Logic Puzzle" Solver (Bayesian Agent): This student is a human-like logician who writes down every possible rule, tests them like a scientist, and builds a perfect theory of how the game works before making a move.
The "Super-Reader" (Large Reasoning Models or LRMs): These are advanced AI systems that have read almost everything written by humans. They haven't played this specific game before, but they can look at the board, think out loud, and figure out the rules almost instantly, just like a human would.

The Experiment
The researchers wanted to know: Which of these students thinks and learns most like a real human?

To find out, they didn't just watch who won the game. They put real humans inside an MRI machine (a giant camera that takes pictures of the brain) while they played these video games. The goal was to see which AI student's "thought process" matched the actual electrical activity inside the human brain.

The Game: A Mystery Box
The games used were like digital mystery boxes. The players didn't know the rules. They had to discover that "if I touch the red square, I die," or "if I push the brown box, it opens a door." The games required figuring out patterns, making guesses, and changing those guesses when they were wrong.

The Results: The "Super-Reader" Wins

Behavior (How they played):
- The Brute Force learners were terrible. They needed to play the game millions of times just to get good. They were slow and clumsy.
- The Logic Puzzle solver was okay, but a bit rigid.
- The Super-Readers (LRMs) were the stars. They learned the rules in a few tries, just like the humans. They figured out the "mystery" of the game at the same speed and with the same efficiency as the people in the MRI machine.
The Brain Scan (The "X-Ray" of Thought):
- This is the most surprising part. When the researchers compared the AI's internal "thoughts" (its digital brain activity) to the human brain scans, the Super-Readers were a perfect match.
- The AI's "thinking" looked almost exactly like the human brain's activity in the visual cortex (where we see things) and the frontal cortex (where we plan things).
- The Brute Force learners? Their "thinking" looked nothing like the human brain. It was like comparing a calculator to a human mind.
- The Super-Readers predicted human brain activity 10 times better than the other AI models.

The "Thinking" vs. "Doing" Surprise
The researchers found something weird about the Super-Readers.

Discovery: When they were figuring out the rules, they were very human-like.
Execution: Once they figured it out, they sometimes got stuck in a loop. If they found a path to win, they would replay that exact same path over and over, even if a shorter path existed. Humans, however, would immediately switch to the shorter, more efficient path once they knew the rules.
The Takeaway: The AI's "brain" (how it represents the game state) is very human-like, but its "muscle memory" (how it executes the plan) is a bit robotic.

Why Does This Matter?
The paper suggests that these "Super-Readers" (Large Reasoning Models) are the first AI systems that don't just act like humans, but actually think like humans.

Usually, AI models are trained specifically for one task (like playing chess). These models, however, learned the game just by reading the rules and looking at the screen, using knowledge they already had from their training. This is exactly how humans learn new things: we use our general knowledge to understand new situations instantly, rather than starting from zero.

In a Nutshell
If you want to build an AI that understands the world the way we do, don't just teach it to memorize patterns (like the Brute Force learner). Instead, give it a vast library of knowledge and let it "think out loud" to solve problems. The paper shows that this approach creates an AI that not only plays games like us but literally lights up its "brain" in the same places we do.

Technical Summary: Reason to Play: Behavioral and Brain Alignment Between Frontier LRMs and Human Game Learners

Problem Statement

Human cognition is distinguished by the ability to rapidly learn abstract structures in novel environments and flexibly deploy this knowledge to guide intelligent action. While modern AI systems, particularly Large Reasoning Models (LRMs), have demonstrated impressive capabilities in static reasoning tasks, it remains unclear whether they can learn and plan interactively in dynamic environments with the same efficiency and flexibility as humans. Existing benchmarks often fail to capture the internal representations of agents during interactive learning, and traditional deep reinforcement learning (RL) agents, while successful in specific domains, often require orders of magnitude more data than humans and lack transferable internal models. Furthermore, computational neuroscience has struggled to find models that predict human brain activity during interactive cognition, as models trained from scratch on specific tasks often produce brittle encoding signals.

This study addresses the question: Can frontier LRMs learn and plan in a way that aligns with both human behavioral trajectories and human brain activity during the discovery and execution of novel video game rules?

Methodology

Dataset and Task

The authors utilized the VGDL-fMRI dataset, which contains concurrent fMRI recordings and gameplay trajectories from 32 healthy adult participants learning to play 12 distinct video games defined by the Video Game Description Language (VGDL).

Task Structure: Participants played 6 games over 6 scanner runs. Each game consisted of 9 levels with increasing complexity, introducing new object types and interaction rules incrementally.
Environment: Games required rule discovery, hypothesis revision, and multi-step planning (e.g., pushing boxes, avoiding hazards, cooperative guidance).
Input Constraints: Models received color-anonymized, symbolic observations (grid positions, colors, inventory) identical to human participants, without privileged semantic labels.

Model Classes

The study evaluated a suite of agents across three paradigms:

Frontier Large Reasoning Models (LRMs): Eight models from the Qwen 3.5 and DeepSeek families (ranging from 9B to 397B parameters). These models were evaluated in a multi-turn dialogue format where they received observations and generated actions. Two conditions were tested:
- Copied-reasoning: The model's hidden chain-of-thought was included in the context for subsequent turns.
- Action-only: Only actions were provided, isolating the contribution of explicit reasoning traces.
Deep Reinforcement Learning (RL) Baselines:
- Model-free: Double DQN (DDQN) with extensive per-game hyperparameter tuning and curriculum-based training.
- Model-based: EfficientZeroV2 (EZV2), which learns a latent dynamics model for planning via Monte Carlo Tree Search.
Symbolic Baseline: The Explore–Model–Plan Agent (EMPA), a theory-based agent that performs Bayesian inference over a hand-coded ontology of game rules.

Evaluation Metrics

Behavioral Alignment:
- Discovery Efficiency: Cumulative steps to the first win per level (measured via Earth Mover's Distance, EMD, against human distributions).
- Capability: Average level reached under a fixed step budget (blocked curriculum requiring two consecutive wins to advance).
Brain Encoding:
- Protocol: LRMs were run as passive observers, feeding them the exact observation stream seen by human participants without generating actions or reasoning traces. Hidden-state features were extracted at the final input token.
- Analysis: Banded ridge regression was used to predict Blood-Oxygen-Level-Dependent (BOLD) responses from model features. Performance was measured as Pearson correlation ( $r$ ) between predicted and observed BOLD activity across six anatomical regions of interest (ROIs): Frontal, Motor, Parietal, Visual, Early Visual, and Striatal.
- Controls: Comparisons included random-initialization controls, trajectory-shuffle controls, and nuisance regressors (button presses, game identity, time).

Key Contributions

Systematic Evaluation of LRMs in Interactive Learning: This is the first systematic evaluation of frontier LRMs on human-like interactive game learning across 12 games and 32 human participants. Unlike prior work using single-step prompting or older LLMs, this study employs multi-step dialogue with explicit reasoning traces.
Behavioral Parity with Humans: The study demonstrates that frontier LRMs match human learning trajectories in both discovery efficiency and capability progression, whereas deep RL baselines require orders of magnitude more experience and plateau early.
Superior Brain Alignment: LRMs predict human brain activity an order of magnitude better than deep RL baselines (DDQN, EfficientZero) and the symbolic EMPA agent across all cortical and subcortical regions. This alignment is robust to permutation controls and is achieved off-the-shelf without fine-tuning or domain-specific tools.
Mechanistic Insight: The study isolates that the brain alignment signal reflects the model's in-context representation of the game state rather than its downstream planning or reasoning processes. This is evidenced by the fact that encoding accuracy remains high even when the model is run as a passive observer without generating reasoning traces.
Resource Release: The authors release an interactive supplementary material containing over 100,000 reasoning traces, all replay data, extracted hidden-state features, and the full analysis pipeline.

Results

Behavioral Performance

Discovery Efficiency: Frontier LRMs clustered tightly around the human distribution of steps-to-first-win. The best model, DeepSeek V4-Pro, achieved a discovery EMD of 0.28, compared to 3.07 for DDQN and 3.22 for EfficientZero. This represents a 5–11 $\times$ reduction in distance to human behavior compared to RL baselines.
Capability: LRMs progressed through the 9-level curriculum at rates comparable to humans (reaching average levels 4–6 within a 1600-step budget), while RL baselines rarely advanced past level 1–2.
Discovery–Execution Gap: A notable divergence was found in execution. While humans significantly shortened their steps after the first win (median 32 steps), LRMs tended to "perseverate," replaying the exact trajectory of their first win even when shorter paths were articulated in their reasoning. This gap diminished with model scale and was largely absent in DeepSeek models.

Brain Encoding

Magnitude of Alignment: LRMs predicted BOLD activity with a Pearson $r$ of approximately 0.07–0.10 in visual and early visual regions, and lower but significant values in frontal, motor, and striatal regions. This is roughly an order of magnitude higher than the best RL baselines (which hovered below $r=0.015$ ).
Training vs. Architecture: Trained LRMs significantly outperformed randomly initialized versions of the same architecture (6.4–6.9 $\times$ improvement), confirming that the alignment stems from learned representations, not architectural priors.
Scale vs. Alignment: Unlike behavioral performance, which generally scaled with model size, brain encoding accuracy did not scale monotonically with parameter count. The Qwen3.5-35B-A3B (MoE) model achieved the highest encoding accuracy, outperforming larger models like DeepSeek V4-Pro.
Representation Specificity: The encoding signal was locked to the moment-to-moment sequence of gameplay. Permuting the temporal order of embeddings or truncating context history significantly reduced encoding accuracy, indicating the signal captures immediate representational structure rather than long-term memory integration.

Significance and Claims

The paper claims to establish Frontier LRMs as compelling computational accounts of human learning and decision-making in complex, naturalistic environments.

Novelty of Alignment: To the authors' knowledge, this is the first joint behavioral and neural alignment between humans and LRMs during interactive gameplay.
Mechanism of Alignment: The results suggest that the rich, transferable representations acquired by LRMs through pre-training on text allow them to approach novel games with a "broad prior knowledge" similar to humans. This contrasts with deep RL agents that must learn from scratch, resulting in brittle representations that fail to align with human brain activity.
Limitations: The authors modestly note that their paradigm makes per-voxel noise ceiling estimation difficult due to the uniqueness of every human trajectory. They also acknowledge that LRMs retain general priors from their training corpora that cannot be fully controlled for.
Future Direction: The paper posits that while current LRMs capture the representational structure of human cognition, they miss the signal associated with active learning and planning itself. Future work requires richer cognitive architectures and active engagement protocols to align models with individual human trajectories.

The study concludes that LRMs offer a path forward for computational neuroscience, providing models that not only solve tasks but also reflect the internal computational states of the human brain during interactive learning.