Turing Laureate Andrew Yao Unveils Algebraic World Models for Lean 4

Modern reasoning models can churn out machine-verified proofs for Riemann zeta functions or Olympiad geometry—provided a human mathematician hands them the exact definitions, axioms, and formal specs. But drop an AI agent into an unfamiliar world with no rulebook, and its formal logic vanishes. The agent must first discover what objects exist, how actions alter state, and what the laws of physics are.

To solve this foundational gap, Turing Award laureate Andrew Chi-Chih Yao, alongside Guancheng Pan, Qixin Tan, and Yang Yuan from Tsinghua and Zhejiang University, have unveiled the Algebraic World Model (AWM). AWM is an end-to-end framework that autonomously converts raw, trial-and-error environmental interactions into typed algebraic representations, category-theoretic abstractions, and kernel-checked Lean 4 mathematical proofs.

The results mark a historic leap in verified agency: AWM propelled Claude Opus 5 from a baseline score of 40.7 to 99.9 on the ARC-AGI-3 benchmark, replayed winning trajectories across all 183 levels in Lean 4, and formally proved whether unknown game levels were mathematically winnable or strictly impossible.

---

The Impossibility Problem: Why LLMs Can't Prove a Negative

When standard AI agents fail at a task, they cannot distinguish whether they simply haven't searched deep enough or whether the goal is mathematically unreachable. In 1736, Leonhard Euler proved the Seven Bridges of Königsberg was impossible not by brute-forcing paths, but by abstracting the islands into vertices, counting odd degrees, and proving no Eulerian path could exist.

Current frontier LLMs cannot perform this abstraction on unseen environments. In baseline tests using Claude Opus 5.5 and GPT-6 Astra on "sealed" game levels (levels where the action budget is programmatically set exactly one unit below the minimum required steps to win), both frontier models failed completely across 12 baseline runs. They either looped until exhaustion, guessed random heuristics, or hallucinatively claimed suboptimal paths (such as estimating a cost of 22 when the true optimum was 19) without mathematical certainty.

In contrast, AWM achieved a 6-for-6 success rate on these sealed levels, synthesizing complete, closed, kernel-checked Lean 4 proofs demonstrating that no legal sequence of actions could ever reach the target state under the allocated budget.

---

Inside AWM: The Three-Stage Mathematical Pipeline

Instead of relying on an opaque neural latent space or massive monte-carlo tree searches, AWM builds an interpretable, verifiable algebraic specification of its environment through three distinct phases:

[ Raw Environment Interaction ]
               │
               ▼
   1. Explore & Prototype
      • Active probing & state logging
      • Executable Python hypothesis generation
      • Continuous refutation against new transitions
               │
               ▼
   2. Structure & Simplify
      • Typed object extraction (States, Actions, Invariants)
      • Category-theoretic projective & inductive limits
      • Rule consolidation (e.g., 11 buttons ➔ 1 permutation law)
               │
               ▼
   3. Formalize & Prove
      • Auto-translation to Lean 4 code
      • Reachability witness verification
      • Invariant-based unreachability proofs

1. Explore & Prototype (Empirical Discovery)

The agent interacts with the black-box environment by dispatching exploratory actions. Rather than training a neural surrogate, the agent generates modular Python functions representing candidate rules and mechanics. Every single observed frame transition is treated as a unit test: whenever an observation contradicts an existing Python concept, the hypothesis is immediately revised or split.

2. Structure & Simplify (Category-Theory Compression)

A fragmented collection of ad-hoc rules quickly explodes into combinatorial chaos. AWM cleanses this representation using category-theory-guided transformations. By computing projective (categorical limits) and inductive limits, AWM unifies redundant domain concepts into shared algebraic structures.

• Case study (Game lp85): In early exploration, the agent drafted eleven separate transition concepts across four distinct button types. The categorical simplification module proved that these eleven concepts were morphisms of a single underlying symmetric group action, consolidating them into one universal typed rule: press_permutes.

3. Formalize & Prove (Lean 4 Kernel Check)

Once the algebraic model stabilizes, AWM compiles the typed objects, state space, and transition monoid into formal Lean 4 definitions.

• For Reachability: The agent generates an action sequence that Lean’s deterministic kernel executes against the formal model from the initial state, verifying a constructive winning witness.
• For Impossibility: The model synthesizes formal state invariants (e.g., parity constraints, Manhattan distance lower bounds, or energy budget conservation laws) and proves by induction that the intersection between reachable states and the winning set is empty.

---

Benchmark Supremacy: Demolishing ARC-AGI-3

The authors benchmarked AWM across all 25 public games (encompassing 183 distinct levels) of ARC-AGI-3—a testbed notorious for demanding abstract spatial, mechanical, and goal-oriented inductive priors.

Metric / Capability	Frontier Baselines (Opus 5.5 / Astra)	AWM + Claude Opus 5	Outcome / Gain
ARC-AGI-3 Official Score	40.7	99.9	+59.2 points
Action Efficiency	100% (Baseline)	43% of human actions	57% fewer steps
Sealed Level Impossibility Proofs	0 / 12 runs	6 / 6 runs verified	100% Kernel-checked
Unseen Layout Transition Accuracy	Degrades over distance	4,160 / 4,160 (100%)	Flawless generalization
Mechanic Mutation Detection	Prone to hallucinated inertia	4,031 / 4,031 (100%)	Instant invalidation

Crucially, Lean 4 replayed and validated the certified clears for all 183 levels. Furthermore, when tested on completely unseen spatial layouts with frozen rules, AWM's models predicted all 4,160 transitions with zero errors. When game developers secretly altered underlying mechanics, the system rejected all 4,031 invalid transitions immediately, avoiding the catastrophic behavioral inertia typical of reinforcement learning policies.

---

The Shift from Neural Latents to Symbolic Truth

For the past three years, the dominant paradigm in world modeling has relied on generative neural simulation—predicting raw pixels or blurry diffusion latents (such as Sora or Genie). While visually rich, neural world models suffer from compounding drift, cannot guarantee consistency over long horizons, and cannot prove safety or reachability bounds.

AWM demonstrates that the future of autonomous agents lies at the intersection of active empirical probing and formal algebraic synthesis:

• Inspectable & Debuggable: Every learned rule exists as human-readable, typed Lean code rather than opaque weight matrices.
• Safety & Guarantees: Mission-critical robotics, financial pipelines, and automated scientific discovery can verify structural impossibilities before burning real-world actions.
• Extreme Sample Efficiency: By leveraging category-theoretic invariants, agents require orders of magnitude fewer environmental samples to master complex operational domains.

By uniting Turing laureate Andrew Yao’s theoretical rigor with modern interactive theorem proving, Algebraic World Models bridge the chasm between raw perception and formal mathematical truth—proving that AI can not only play in an unknown world, but write its fundamental laws.