Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Reason to Play: Behavioral and Brain Alignment Between Frontier LRMs and Human Game Learners
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Compares frontier reasoning models (Qwen3.5 and DeepSeek V3.2/V4 families) with deep RL, EMPA and human learners on VGDL games, showing LRMs match human discovery behavior and predict fMRI BOLD an order of magnitude better than RL baselines.
Problem
Whether modern AI systems can learn and plan interactively in novel video games like humans, assessed jointly through behavior and concurrent brain activity during game learning.
Approach
Evaluates frontier LRMs in a multi-turn dialogue gameplay setup on VGDL games and uses their in-context hidden-state representations (extracted from passively observed human trajectories) to predict human fMRI BOLD, comparing against DDQN, EfficientZero, and EMPA.
Key results
- DeepSeek V4-Pro achieves discovery EMD of 0.28 to human distribution vs 3.07 for DDQN and 3.22 for EfficientZero; EMPA reaches 0.49.
- LRM representations predict BOLD about an order of magnitude better than RL baselines across six region groups, with 6.4-6.9x above random-initialization controls (paired t > 14, p < 10^-11; n=21).
- LRMs show a discovery-execution gap: median execution steps of 45-88 versus human 32, driven by perseveration on replayed trajectories.
Surprising findings
- Encoding accuracy does not scale monotonically with parameter count; Qwen3.5-35B-A3B outperforms larger DeepSeek models on brain encoding.
- Prompt-supplied knowledge (minimal vs elaborate suggestion levels) and long context truncation do not change brain encoding accuracy.
Limitations
- Per-voxel noise ceilings are hard to estimate because every gameplay trajectory is unique, so regional differences largely reflect SNR.
- LRMs carry opaque pretraining priors that cannot be controlled, and baselines receive different forms of game information (EMPA has ground-truth ontology).
- Brain encoding reflects in-context representations of the current state rather than learning or planning; LRMs show perseveration in execution.
Experimental setup
Datasets: VGDL-fMRI dataset (Tomov et al.)
Benchmarks: VGDL-fMRI human gameplay and BOLD data
Metrics: log-space Earth Mover's Distance solve rate Pearson r encoding accuracy capability progression
Baselines: DDQN EfficientZeroV2 EMPA HRR random-initialization LRM
Builds on: Double DQN EfficientZeroV2 EMPA VGDL banded ridge encoding
Closest papers by meaning
- Maia-2: A Unified Model for Human-AI Alignment in Chess (NeurIPS 2024 · similarity 0.87)
- Human-Aligned Chess With a Bit of Search (ICLR 2025 · similarity 0.87)
- Beyond Single Stationary Policies: Meta-Task Players as Naturally Superior Collaborators (NeurIPS 2024 · similarity 0.87)
- An Evolutionary Perspective on AI Alignment (Student Abstract) (AAAI 2025 · similarity 0.86)
- The Algorithm Is Not the Behavior: Learned Priors Override Look-Ahead in a Chess-Playing Neural Network (NeurIPS 2026 · similarity 0.86)
- Large Language Models Think Too Fast To Explore Effectively (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.