A credible benchmark for agents that have to plan.
WorldModel Gym measures long-horizon planning under sparse rewards and partial observability — the regime where reactive policies stall and a learned model of the world starts to matter. It is built to be reproducible, honestly measured, and read at a glance.
Plan over a learned model
An agent imagines rollouts through a world model, then commits to the path most likely to reach a goal it usually cannot see.
Why this benchmark
The hard part is what you cannot see.
Most benchmarks reward fast reactions. WorldModel Gym isolates the three pressures that force an agent to actually model the world.
Long-horizon planning
Goals sit dozens of dependent steps away. An agent must look ahead through a learned model of the world, not react one frame at a time.
Sparse rewards
Reward arrives only at a terminal goal event. There is no dense shaping to climb, so credit assignment has to survive long stretches of silence.
Partial observability
A limited field of view hides most of the state. The agent has to remember, infer, and imagine what it cannot currently see.
Scientific methodology
Reproducible by construction, honestly measured.
A benchmark is only as credible as its bookkeeping. Three commitments keep the numbers trustworthy.
Reproducible tracks
Seeded test, train, and continual
Every environment ships fixed seeds across three tracks. Test seeds are held out from training; the continual track replays a shifting task stream so non-stationary adaptation is measured, not assumed.
- Deterministic episode generation from a track seed
- Held-out test seeds, disjoint from train
- Continual track for drift and forgetting
Honest metrics
Success from terminal goal events
Success counts only true terminal goal events from the trace — no proxy heuristics. Aggregate numbers carry bootstrap confidence intervals, and model quality is reported as true k-step fidelity, not a single-step proxy.
- Success = terminal goal event in the episode trace
- Bootstrap CIs on every aggregate metric
- True k-step world-model fidelity, not 1-step
Credible baselines
From BFS oracle to real PPO
Each track is anchored by reference agents spanning the difficulty spectrum, so a new submission is read against a meaningful floor and ceiling rather than a single cherry-picked number.
- BFS oracle — shortest-path upper bound
- Value-bootstrapped MCTS and MPC-CEM planners
- A real PPO learner, not a placeholder
Reference baselines
A meaningful floor and ceiling.
Every track is anchored from a shortest-path oracle down to a real model-free learner, so a new run is read against context — not in a vacuum.
Breadth-first shortest path on the true state — the practical ceiling for each seed.
Tree search over the learned model, bootstrapped from a learned value estimate.
Cross-entropy method rolling out action sequences inside the world model.
A genuine on-policy RL baseline — the honest model-free reference point.
Architecture at a glance
One pipeline, from environment to public run.
A Python research core feeds an evaluation server; this dashboard renders the result as a benchmark you can share with a URL.
Environments
MemoryMaze, SwitchQuest, and CraftLite — procedural, sparse-reward, partially observed.
Planners + world models
MCTS, MPC-CEM, and trajectory sampling over deterministic, stochastic, and ensemble latents.
Evaluation server
FastAPI runs CRUD, artifact uploads, leaderboard queries, and trace/metrics downloads.
Dashboard
This Next.js App Router surface — tasks, leaderboards, run viewer, and compare.
Get involved
Use it, extend it, cite it.
Submit a run from the browser, the API, or the CLI. If WorldModel Gym helps your work, a citation keeps the benchmark honest and traceable.
Cite this benchmark
Reference WorldModel Gym in papers, READMEs, or model cards.
@software{worldmodelgym2026,
author = {Saikia, Birajit},
title = {{WorldModel Gym: A Benchmark for Long-Horizon Planning under Sparse Rewards and Partial Observability}},
year = {2026},
url = {https://world-model-gym.vercel.app},
note = {Reproducible seeded test/train/continual tracks with honest success, bootstrap CIs, and k-step model-fidelity metrics. Version 0.1.0.},
}