Building Roller Coaster Tycoon Bench: Can Decision Models Manage a Theme Park?
Blog post from Vals
RollerCoasterTycoonBench evaluates nine decision models by having each manage a RollerCoaster Tycoon 2 park for 96 weekly turns across three runs, aiming to finish with at least 1,000 guests and a rating of 600; scores combine goal attainment and company value. Decider V1.3 achieved the highest average score, 71 out of 100, and reached the guest target in all three runs, while Decider V1.2 and GPT-6 Luna Decisions each did so once. The runs revealed varied management patterns: Decider V1.3 borrowed early and quickly invested, GPT-6 Luna Decisions adjusted prices based on guest feedback, and some models wasted affordable building opportunities, neglected maintenance, or accumulated unsustainable debt. Models’ stated confidence did not meaningfully track performance, and seven of nine averaged more guests when construction prices were included in the park description, though the small number of runs leaves those differences uncertain. Experiments also found that removing the stated role and objective changed several models’ choices, while option order and other harness rules had more limited effects. The benchmark uses OpenRCT2 with a text-only description of the park and publicly records each decision, providing a way to study how models plan and act over time rather than answer isolated questions.
| Trend | Post Mentions | Total Month Mentions | Posts | Companies | MoM |
|---|---|---|---|---|---|
| Jev | 7 | 409 | 40 | 22 | -68% |
Use this post, company, and trend context to find content marketing opportunities, perform competitive analysis, or address product feature gaps via the Plushcap MCP server or the Plushcap API.