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

Abstract: 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 a similar way? We study this question using a dataset of complex human gameplay with concurrent fMRI recordings, in which participants learn novel video games that require rule discovery, hypothesis revision, and multi-step planning. We jointly evaluate models by their ability to play the games, match human learning behavior, and predict brain activity during the same task, comparing a suite of frontier Large Reasoning Models (LRMs) against model-free and model-based deep reinforcement learning agents and a Bayesian theory-based agent. We find that frontier LRMs most closely match human behavioral patterns during game discovery and predict brain activity an order of magnitude better than both reinforcement learning alternatives across cortical and subcortical regions, with effects robust to permutation controls. Through targeted manipulations, we further show that brain alignment reflects the model's in-context representation of the game state rather than its downstream planning or reasoning. Our results establish LRMs as compelling computational accounts of human learning and decision making in complex, naturalistic environments. Project page with interactive replays: https://botcs.github.io/reason-to-play/

• The paper evaluates the alignment of frontier Large Reasoning Models with human-like learning and cognitive representations in game environments.
• Using VGDL tasks and fMRI data, the study examines how LRMs compare to human benchmarks and outperform model-free deep RL agents.
• Findings show LRMs exhibit human-like traits, but fall short in tasks requiring advanced reasoning like multi-step subgoaling.

Behavioral and Neural Alignment of Frontier Large Reasoning Models With Human Game Learning

Introduction

The capacity for rapid, abstract generalization in novel environments is a hallmark of human cognition. The study assesses whether frontier Large Reasoning Models (LRMs)—LLMs post-trained to produce explicit reasoning traces—exhibit human-like learning and internal representations in complex, rule-inductive game environments. Using interactive Video Game Description Language (VGDL) tasks coupled with fMRI from 32 human participants, the authors provide a rigorous, multi-modal evaluation of behavioral and brain alignment between a suite of LRMs and humans, juxtaposed against model-free/model-based deep RL agents and a Bayesian symbolic agent (EMPA).

Experimental Design and Evaluation

The VGDL-fMRI dataset presents a highly controlled sandbox for rule discovery, hypothesis testing, and multi-step planning. The evaluation covers eight cutting-edge LRMs (Qwen 3.5 and DeepSeek) in two operational regimes—a copied-reasoning (chain-of-thought in context) and action-only (no rationale) dialogue protocol—ensuring all task-relevant information remains in-context and not encoded in the weights via per-game training.

Model representations are evaluated along three axes:

1. Behavioral alignment: Steps to first win (discovery efficiency), solve rates, capability progression across blocked curricula.
2. Neural encoding: Predictive power of model hidden states for fMRI BOLD responses in human brain regions during gameplay.
3. Reasoning interpretability: Analysis of rationale content and dynamics throughout play.

Key Findings

Behavioral Alignment With Human Learning

Frontier LRMs closely track human behavioral distributions in cumulative steps to first win, with DeepSeek V4-Pro and Qwen3.5-35B-A3B exhibiting Earth Mover's Distance (EMD) scores of 0.28 and 0.40, respectively (humans ~0.00; DDQN/EfficientZero ~3.1). This reflects an order-of-magnitude improvement in efficient, human-like knowledge acquisition over both model-free and model-based RL. EMPA, with access to a privileged ontology, is closer to humans than deep RL but still lags behind the LRMs.

Solve rates also approach human performance, with top LRMs solving 55–65% of level-instances. However, all models fall short on games requiring multi-step subgoaling and non-trivial cooperative reasoning (HELPER, LEMMINGS), marking a limitation in current LLM architectures for compositional/relational reasoning under partial observability.

A notable discrepancy is the discovery-execution gap: unlike humans, LRMs display a tendency to rigidly replay successful trajectories for subsequent wins without compressing their policy, reflecting limited route optimization and generalization in execution phases.

Neural Encoding and Representational Alignment

The central claim is that LRM-derived hidden state features predict human BOLD responses an order of magnitude above RL and symbolic baselines across all cortical and subcortical ROIs. Best-performing LRMs achieve Pearson $r$ values in the range $0.07–0.10$ in visual cortex (DDQN/EfficientZero $<0.015$; HRR $\sim0.02$), and demonstrate robust encoding performance in parietal, frontal, motor, and striatal regions as well.

Importantly, randomly-initialized versions of LRM architectures are non-predictive (encoding accuracy matches RL baselines), confirming that the signal reflects learned representations, not architectural priors. Permutation controls further show the representational signal is temporally locked to human-perceived game state, not model action or reasoning output.

Critically, LRM–brain representational alignment does not scale monotonically with parameter count, diverging from behavioral capability scaling. Mixture-of-Experts configurations (e.g., Qwen3.5-35B-A3B) outperform larger dense models in neural prediction tasks, decoupling model size from internal alignment with human representations. This resonates with findings in brain-like vision models, suggesting an architectural and training-regime dependency for optimal brain alignment.

Qualitative Reasoning Analysis

Rationale traces in the copied-reasoning regime provide step-wise correspondence between hypothesis formation, causal verification, and adaptive planning, paralleling human think-aloud protocols. Explicit chain-of-thought is essential for behavioral and neural alignment: stripping it (action-only) produces a uniform collapse in both solve rate and neural predictivity, emphasizing the role of in-context reasoning in flexible adaptation and representational richness.

Implications

Theoretical Implications

The work substantiates LRMs as computational models of human-like abstraction and in-context learning in structured, interactive domains. Their match to both human behavior and representational geometry in cortex—without task-specific fine-tuning—supports the hypothesis that pretrained text-based models acquire transferable cognitive priors paralleling those used by humans.

The decoupling between execution compression (route optimization) and discovery (rule induction) exposes a dissociation in LLMs between episodic knowledge re-use and systematic generalization, pointing toward future work integrating explicit model-based planning or compositional memory architectures into reasoning models.

Finally, the divergence between behavioral capability and brain-alignment scaling impels a nuanced view of model selection: optimal alignment with human representations may arise at intermediate scale, architecture-dependent, and based on post-training objectives, rather than simply maximizing size or held-out task performance.

Practical Implications and Prospectus

The results position off-the-shelf LRMs as strong candidates for simulating or probing human cognitive strategies in open-ended tasks, with potential application in cognitive neuroscience—as computational probes of representation—education, and interactive AI systems requiring robust generalization and adaptation from sparse data.

Future research must address three limitations:

• Incorporating active engagement protocols (beyond passive observation) to align models and humans at the trajectory level, potentially via closed-loop harnesses and individualized training curricula.
• Disentangling which cognitive operations (e.g., planning, exploration-exploitation, theory revision) underlie the observed brain-alignment.
• Expanding the evaluation to include adversarial rule changes and more complex compositional tasks to stress-test model adaptability and meta-learning.

Conclusion

This study presents a comprehensive joint behavioral and neural assessment of frontier Large Reasoning Models in interactive, inductive environments. LRMs exhibit human-like learning trajectories and achieve robust brain representational alignment, far surpassing deep RL and symbolic models, without access to privileged ontologies or environment-specific pretraining. The findings affirm the promise of large-scale reasoning models in computational cognitive science, delineate the boundaries of current architectures, and lay the groundwork for future integration of active cognition, compositional reasoning, and brain-constrained AI systems.

---

Reference:

"Reason to Play: Behavioral and Brain Alignment Between Frontier LRMs and Human Game Learners" (2605.08019)