Crown of Defeat and Fire of Miracles: An Apocalypse of AlphaProof Nexus's 95.7% 'Failure Rate'

*Translated and adapted from a zhichai.net forum analysis.*

Introduction: A Strange Celebration in Science

Google DeepMind released an AI system called AlphaProof Nexus and reported a 95.7% "failure rate": when attempting 350 open mathematical problems, it solved only 9.

If a student scored 2.5% on a final exam, they would fail outright. Yet this sky-high failure rate has top mathematicians and Google's research team as excited as children. The reason is simple: the 341 unsolved problems are world-class open questions that have stumped humanity for decades — while the 9 that were solved come from the ultimate legacy left to humanity by the legendary mathematician Paul Erdős.

A top mathematician might solve one or two such problems in a lifetime. This AI knocked out 9 in a matter of days, for a few hundred dollars of compute.

This was not a crushing defeat. It was a major victory at the boundary of abstract reasoning.

1. Ending the Hallucination Dream: The Lean Formal Sandbox

When GPT-4 or Claude work on math problems, they often produce text that looks rigorous but contains hidden logical gaps. This is the fatal weakness of large language models — hallucination.

AlphaProof Nexus abandons natural language entirely. Its proofs are all written in the Lean language.

> Lean (interactive theorem prover): A proof-assistant language and computer-assisted mathematics tool based on type theory. It translates mathematical theorems and proofs into strictly machine-checkable code. The Lean compiler validates every derivation step against the axiomatic system at the deepest level, tolerating no logical gaps. > > Formal verification: Using rigorous mathematical methods to prove the correctness of a system or algorithm. In Lean, a proof is only accepted as true after passing the compiler's axiom-dependency checks — completely eliminating human oversight lapses and AI confabulation.

Inside Lean's sandbox there is no compromise, no "probably." A proof has only two possible outcomes: "compiles" or "error." This severs the possibility of AI hallucination at the physical layer. By translating natural-language conjectures into Lean statements, Nexus aligns itself with mathematical truth line by line, under the cold supervision of the Lean compiler.

2. Playing Against Itself: Elo Ratings and an Agent Feedback Network

How can a few hundred dollars of compute outperform a human mathematician's lifetime of effort? The secret is Nexus's self-play agentic adversarial network:

1. Multi-agent advance: The system is not a single neural network thinking. It deploys a large batch of "Prover" agents, each seeking a breakthrough. 2. The cold Judge agent: When a Prover generates a Lean proof, the Judge agent feeds it to the Lean compiler. If compilation fails, the compiler's error messages (e.g., "type mismatch," "invalid axiom reference") are converted into feedback vectors and returned to the Prover. 3. Elo rating system: To keep the algorithm out of dead ends, Nexus uses a chess-style Elo rating mechanism. Provers producing valid proof steps — or shrinking the proof-tree search space — gain Elo points and receive more compute in later rounds; Provers making repeated blunders are demoted and eliminated.

This self-evolution under Lean's rules means the system needs no curated corpus of "correct proofs." Through pure probabilistic search and self-play across millions of failures, it carves out its own path to truth.

3. Dimensionality Reduction: From Brute Force to Probabilistic Pruning

Brute-force enumeration of infinite axiom combinations would exceed the compute of every machine on Earth combined.

Nexus instead converts proof search into heuristic Monte Carlo Tree Search (MCTS):

• The neural network observes the current Lean state and outputs a probability distribution \(\pi(a|s)\) over the next most useful theorems and algebraic transformations;
• The algorithm explores deeply only along the highest-probability branches, pruning the infinite search space into narrow, precise channels.

The arithmetic is stunning:

\[\text{compute cost} \propto \text{search tree depth} \times \text{branching factor}\]

By compressing the branching factor by several orders of magnitude, proofs that would have required supercomputers running for years were squeezed into a few hundred dollars of cloud CPU/TPU billing.

4. The Final Echo: Algebraic Geometry and a New Scientific Paradigm

Beyond the 9 Erdős problems, AlphaProof Nexus reportedly:

• Proved 44 long-unsolved conjectures cataloged in OEIS (the On-Line Encyclopedia of Integer Sequences);
• Solved an open classification problem of geometric structures in algebraic geometry that had stood for 15 years.

This marks a paradigm shift in scientific research. Future mathematicians will not need to spend years verifying tedious details on blackboards. They will formulate grand conjectures and architectures, while machines like AlphaProof Nexus — cheaply and at high speed — carve out paths to truth in a cold formal sandbox.

5. Academic Appendix: Citations and Sources

1. *Automated Theorem Proving and Research-level Mathematics with AlphaProof.* Google DeepMind, 2026.

• Systematically describes the AlphaProof Nexus architecture and demonstrates proving 9 open Erdős problems autonomously in the Lean sandbox via reinforcement learning and self-play, without human-curated high-quality proof corpora.
2. *The Lean theorem prover: A formal foundation for computer-assisted mathematics.* (Moura et al., Journal of Automated Reasoning, 2015).
• Defines the dependent type theory kernel of Lean and its kernel-checking mechanism, establishing the modern standard for machine-verified, computer-assisted mathematics.
3. *Self-play and agentic feedback loops in reinforcement learning for proof search.* (ASPLOS, 2025).
• Shows that introducing Elo rating and compiler error feedback loops into sparsely-rewarded proof search trees can raise invalid-branch pruning rates by over 70%.

*Note: This is a translation of a community analysis; the cited papers and figures are as stated in the original post.*