Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Mini Amusement Parks (MAPs): A Testbed for Modelling Business Decisions
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.
Introduces Mini Amusement Parks (MAPs), a business-management simulator benchmark with expert human baselines, showing frontier LLM agents, PPO, and WALL-E world-model MPC lag far behind experts on open-ended optimization, long-horizon planning, active world-model learning, spatial reasoning, and stochasticity.
Problem
Existing human-AI benchmarks do not jointly test open-ended optimization, sample-efficient world-model learning, long-horizon planning, spatial reasoning, and stochastic transitions in a grounded business-style decision-making setting.
Approach
An amusement-park management simulator with stochastic guest behaviour, a sandbox mode with auxiliary exploration actions for active world-model learning, easy and medium difficulties, and expert human leaderboard scores as a ceiling; agents are evaluated with ReAct, PPO, VLM variants, a spatial heuristic, and WALL-E-based MPC.
Key results
- GPT-5 achieves 8.80% of the human upper bound on easy mode, the best of tested systems; experts outperform by 11.4x on easy and 15.3x on medium.
- GPT-5 drops from 8.80% (easy) to 4.44% (medium) as the horizon doubles and research is introduced.
- Oracle world-model MPC gives roughly 3.5x improvement over ReAct for GPT-5 Nano on easy mode, while WALL-E MPC degrades performance.
Surprising findings
- Sandbox learning provides little to no benefit and often lowers scores; learned notes often parrot the manual or are overly specific.
- Adding the park image to a VLM reduces performance across all three models, while a simple spatial heuristic consistently improves it.
Limitations
- Human upper bound is from unlimited-resource play, not a directly comparable baseline.
- Only easy and medium modes and three test layouts; generalization may not be fully characterized.
- RL baselines are intentionally simple; hierarchical or model-based RL not evaluated.
Experimental setup
Datasets: MAPs training layouts MAPs sandbox experience
Benchmarks: MAPs
Metrics: final park value normalized to human upper bound (%) runtime (s) cost (USD) coefficient of variation
Closest papers by meaning
- EcoGym: Evaluating LLMs for Long-Horizon Plan-and-Execute in Interactive Economies (NeurIPS 2026 · similarity 0.86)
- Agentick: A Unified Benchmark for General Sequential Decision-Making Agents (NeurIPS 2026 · similarity 0.83)
- Wide-Horizon Thinking and Simulation-Based Evaluation for Real-World LLM Planning with Multifaceted Constraints (NeurIPS 2025 · similarity 0.83)
- Configurable Mirror Descent: Towards a Unification of Decision Making (ICML 2024 · similarity 0.82)
- Strategic Planning: A Top-Down Approach to Option Generation (ICML 2025 · similarity 0.82)
- Scaling Autonomous Agents via Automatic Reward Modeling And Planning (ICLR 2025 · similarity 0.82)
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.