An agent learns the rules, then proves a sealed level unwinnable: AWM returns Lean-checked proofs where twelve baseline runs return none
Synopsis
In game environments whose rules and goals are not disclosed, AWM turns hypotheses gathered through interaction into an explicit algebraic world model and then produces kernel-checked Lean 4 proofs of reachability and impossibility. On two sealed levels whose budgets sit below the minimum winning cost, all six AWM runs yield closed, kernel-checked proofs excluding every legal winning sequence within budget, while twelve baseline runs of Claude Opus 5.5 and GPT-6 Astra produce no such proofs. Across 25 public ARC-AGI-3 games, these models let Lean replay the winning trajectories of all 183 levels.
Interpretation
An agent can first learn an explicit mathematical model of an environment whose rules it does not know, then use that model to prove a goal cannot be reached within a given budget, rather than inferring it from repeated failure. Earlier formal-mathematics progress typically rests on predefined definitions and formal specifications, so the objects, action effects and success conditions of the environment need not themselves be learned. On two sealed levels whose budgets were set below the minimum winning cost, all six AWM runs produce closed, kernel-checked impossibility proofs excluding every legal winning sequence within budget, while Claude Opus 5.5 and GPT-6 Astra produce no such proofs across twelve baseline runs.
The model is not a black box: it consists of typed objects, named rules and a goal predicate, so rules can be inspected one by one and proof dependencies show which rules a conclusion rests on. Earlier code world models can reproduce observed transitions, but reproducing sampled transitions does not guarantee that a key rule has been captured. Across the 25 public games, AWM records 742 executable concepts and withdraws 143, organizing the rest into 170 object types and 283 rules, a median of 6 types and 10 rules per game; Lean replays the winning trajectory of all 183 levels from the initial state and verifies the final state.
The learned rules also apply to screens never seen: when a layout changes they keep predicting the next frame, and when a mechanic changes they judge the new transitions inconsistent. Concepts written during exploration cover only the transitions actually traversed and accept noticeably fewer valid transitions on unseen screens. On ls20, concepts from exploration accept 67%–70% of valid transitions on unseen screens, while the frozen rules after structure and simplification accept 99.47%–99.95%; on the two ft09 sets with a changed mechanic, the frozen rules reject all 4,031 transitions.
Perspective
The demonstrated scope is ARC-like environments: discrete actions, immediate and observable feedback, and learned rules that can be checked transition by transition against recorded data. The authors list richer hidden state, delayed feedback and stochastic dynamics as directions still to be tested.
Lean guarantees that a proof holds under the learned model and its explicit premises; whether those premises capture the environment is tested by interaction. On the two sealed levels, two initial runs reported bounds of 20 and 47 because the learned rules omitted cheaper routes the engine allows. Transferring model-level impossibility proofs to the games themselves was verified against engine-generated transitions in five of the six runs.
