Introduction
What is WorldModel Gym
An end-to-end benchmark platform for long-horizon planning agents under sparse rewards and partial observability.
WorldModel Gym is an end-to-end benchmark for long-horizon planning agents. It pairs reproducible environments with a FastAPI submission service and this Next.js leaderboard, so an agent run can travel from evaluation harness to public dashboard through one stable API.
The platform is built around three ideas: tasks should test procedural generalization (not memorized seeds), submissions should be reproducible and idempotent, and the metrics that rank agents should report uncertainty, not just point estimates.
How a run flows
- Create a run row with
POST /api/runs— this returns a run id. - Upload artifacts (
metrics.json, optionaltrace.jsonlandconfig.json) toPOST /api/runs/{run_id}/upload. - The server scores the run and surfaces it on the leaderboard by track.
- Browse the run, inspect its planner trace, and compare agents from the dashboard.
Same-origin by design
The browser never calls the API cross-origin. The web app proxies every request through /api/proxy/*, which forwards to the upstream FastAPI service and copies back the x-request-id for traceability.
What you can do
- Create benchmark runs and upload metrics, traces, and config artifacts.
- Inspect public leaderboard data filtered by track, environment, and agent.
- Browse tasks and benchmark context, including planner traces and grid worlds.
- Verify deployment health with readiness, liveness, and metrics endpoints.
New here? Jump straight to the Quickstart to submit your first run, or read Concepts to understand tracks, environments, and metrics.