Observe
Receive a few anonymous symbols. The latent state and rulebook remain hidden.
AGI benchmark · research proposal
A world with
no manual.
Reworld asks a machine to do something closer to science than question answering: enter an unfamiliar world, run experiments, infer its hidden causal structure, and learn to control it.
Current status: specification and falsification design. No benchmark score is live yet.
LATENT
Wt
seed 7A·31
relation
memory
effect
SURFACE / 12 BITS
01One active worldShared laws, isolated runs
02Fixed embodimentThe interface never grows
03Generated causalityNo LLM-authored rulebook
04Operational proofUnderstanding shown by control
The question
Can an agent learn
science from scratch?
Most benchmarks present tasks in familiar representations and ask for an answer. Reworld presents a continuing causal system and asks the agent to discover what questions are worth asking.
The surface is intentionally simple and stable. Almost everything below it—the entities, relations, operators, memory, delays, and layered interactions—can change with each world.
Experiment
Manipulate fixed controls, reset instances, and design interventions that separate hypotheses.
Model
Compress the evidence into a predictive causal model—not necessarily the simulator’s own representation.
Control
Demonstrate understanding by reliably producing progressively deeper target observations.
A permanent body
The surface stays
small and legible.
Every world uses the same narrow instrument panel: a handful of manipulation slots, observation slots, a step operation, reset, target, and visible transcript.
- ✓ Identical symbolic data for humans and agents
- ✓ Stable controls; changing causal meaning
- ✓ No latent state, labels, or hints leak through the renderer
The panel shown here is an illustrative toy, not a generated Reworld instance.
REWORLD / INTERFACE CONCEPT
INSTRUMENT 04×06
deterministic toy
TARGET
◆●▲× 2 steps
MANIPULATIONinput batch
OBSERVATIONoutput batch
STEP 000
VISIBLE TRANSCRIPTlatest first
- 000
M [0 0 0 0]O [0 0 0 0 0 0]
Below the panel
A thin surface over
causal depth.
Higher layers are not decorative labels. They count only if they form grounded, low-error abstractions that improve prediction, compression, or control.
MANIPULATION
mt
↓
LnObservable surfaceGenerated transducers · narrow bandwidth
grounded abstraction
L2Relations among persistent motifsPredictive equivalence · bounded error
grounded abstraction
L1Persistent motifs and historiesMemory · aggregate events · context
generated dynamics
L0Anonymous hypergraph rewritingFinite symbols · typed relations · local rules
↑
ot+1
OBSERVATION
No authored puzzle book
Worlds are selected,
not written.
The proposal does not use an LLM to invent mechanics. It uses a bounded formal grammar, seeded mutation, deterministic validation, simulation, and a suite of search and learning systems.
ConstitutionTyped state, rule grammar, bounds, execution semantics
VariationSeeded mutation of symbols, relations, rules, and transducers
ValidityClosure, reference integrity, termination, bounded work
BehaviorObservable complexity, causal reach, predictability
FrontierNeither trivial nor noise; learning beats blind search
retain(W)
=
validobservablelearnablenovelbounded
One public number
Score the frontier
an agent reaches.
Each world presents an ordered ladder of experimentally certified targets. The public score is the deepest calibrated level controlled before that world rotates.
median( St, St−1, St−2 )
WORLD 4218.7
WORLD 4116.2
WORLD 4021.4
Illustrative values. Reworld has no live leaderboard yet.
Deepest calibrated control
A level counts only when an observation-contingent policy succeeds across fresh instances and declared side constraints.
Efficiency becomes depth
With one fixed interaction window and resource envelope, inefficient learning already means reaching less of the world.
Continual adaptation
One exceptional or weak world cannot dominate. World-specific solutions age out as the laws rotate.
Fresh certification
Discovery happens under a budget. Success is then tested on fresh, post-commitment instances.
Hosted autonomy
Agent artifacts are hashed before reveal, isolated, resource-bounded, and run without network or human intervention.
Observable calibration
Difficulty comes from behavioral response curves—not latent rule count or a claim about hidden bits.
Public commitments
World selection, rules, instances, and level definitions are committed before activation and auditable afterward.
Build the test before the platform
A falsifiable
research programme.
The public benchmark is downstream of a smaller question: can formal worlds be found where feedback-driven learning reliably beats blind sequence optimization?
Specify steps, seeds, budgets, metrics, baseline suites, and falsification thresholds.
NOWBuild tiny, exactly analyzable worlds with a hand-parameterized compositional generator.
Compare feedback policies, model learners, and strong open-loop search under equal budgets.
Add quality-diversity evolution only if the hand-built generator produces a viable difficulty band.
Retain abstractions only when they measurably improve prediction, compression, or control.
Build the fixed shell, hosted evaluator, commitments, rotation, and longitudinal validation.
GO / NO-GO
The project advances only when its claims survive contact with baselines.
Feedback necessity
Observation-contingent policies succeed where strong open-loop optimizers fail.
Learnability
Prediction and control improve materially with evidence rather than remaining noise.
Observable growth
Accepted world complexity creates new predictive distinctions at the surface.
Transfer
Methods learned on earlier worlds accelerate adaptation to held-out generator families.
Calibration
Independent anchor systems induce a stable enough ordering for one depth score.
Evolutionary value
Evolution produces a better frontier than the hand-parameterized generator.
Research context
Adjacent work,
a different synthesis.
SystemCore ideaReworld’s intended distinction
ARC-AGI-3Interactive reasoning across authored grid gamesOne rotating generated law system behind a much narrower surface
DeepMind AlchemyInfer a resampled latent causal structure through experimentGenerate more of the ontology, operators, and layered dynamics
XLand / AdARapid adaptation across a vast procedural 3D task spaceRemove familiar physics and isolate a fixed low-bandwidth embodiment
POET / PAIREDGenerate curricula and environments at a learning frontierApply that pressure to causal law systems and observable discovery
Reworld does not yet establish superiority to these systems. Its proposed contribution is the combination of generated causal laws, fixed narrow embodiment, operational control targets, one rotating public world, and explicit falsification gates.
Clarifications
What Reworld
is—and is not.
Is Reworld a game?
The surface borrows the legibility of a game, but the purpose is AGI evaluation and research. Humans may inspect the interface and provide reference data; official scored runs are autonomous agent executions.
Are the worlds generated by an LLM?
No. Candidate worlds come from a formal typed grammar, seeded mutation, deterministic validation, simulation, and selection against observable and learnability criteria.
Does a larger hidden world automatically make a harder benchmark?
No. Only the induced action/observation process matters. Hidden complexity that never creates a learnable surface distinction is discarded.
Why can the hidden starting context vary?
The visible panel, world laws, and task stay the same. Small hidden instance variation ensures that a feedback policy—not one memorized action string—is required to demonstrate control.
Can brute force be eliminated?
No benchmark is independent of compute. Reworld’s target is narrower: under a fixed execution envelope, model-building and informative experiments should outperform blind enumeration by a large, measured margin.
Is the world infinitely complex?
Each world remains finite and bounded so validity can be checked. The intended frontier is practically unbounded relative to current systems, not mathematically infinite.
Does a Reworld score equal progress toward AGI?
No. It would measure active system identification and control under severe partial observation on a held-out generator distribution. Broader AGI claims require demonstrated transfer beyond that distribution.
The proposition
Do not tell the agent
how the world works.
Give it a body. Give it observations. Let it intervene. Then measure how much causal depth it can make intelligible.