PKU AI Disproves Anderson Conjecture: China's First Autonomous Math Breakthrough

The Groundbreaking Achievement at Peking University

Peking University (PKU), one of China's premier institutions, has achieved a historic milestone in artificial intelligence applied to pure mathematics. The university's AI4Math team has developed an autonomous AI framework that disproved the Anderson Conjecture, an open problem in commutative algebra that had puzzled mathematicians since 2014. This marks the first time a domestic Chinese AI system has independently resolved a research-level mathematical conjecture and formally verified the solution, showcasing PKU's leadership in higher education innovation. 70 68

The breakthrough not only advances commutative algebra but also demonstrates how AI can bridge disciplinary gaps, accelerating discoveries that might take human researchers years. At PKU's Beijing International Center for Mathematical Research (BICMR), this success highlights the synergy between traditional mathematical expertise and cutting-edge machine learning.

Understanding the Anderson Conjecture in Commutative Algebra

Proposed by D.D. Anderson from the University of Iowa, the Anderson Conjecture addresses properties of Noetherian local rings—a fundamental structure in commutative algebra used to study geometric objects locally. A Noetherian local ring (R, m) is quasi-complete if for every decreasing sequence of ideals {A_n} and each k ≥ 1, there exists s_k such that A_{s_k} ⊆ (∩ A_n) + m^k. The weak version assumes ∩ A_n = 0, simplifying to A_{s_k} ⊆ m^k.

The conjecture posited that weak quasi-completeness implies full quasi-completeness for such rings. Proving or disproving this required deep knowledge across subfields like integral domains and completions, making it challenging for individual mathematicians. 69

PKU's AI4Math Team: A Multidisciplinary Powerhouse

Led by Boya Distinguished Professor Bin Dong at BICMR, the AI4Math team formed organically in 2023 from researchers in algebra, number theory, optimization, and machine learning. Key members include Academician Ruochuan Liu, Dean of PKU's School of Mathematical Sciences; Liang Xiao, an IMO gold medalist and algebra expert; and Zaiwen Wen, an operations research specialist. Their collaboration, supported by IQuest Research Institute and New Cornerstone Science Foundation, exemplifies PKU's vibrant ecosystem for young talent. 70

"This result demonstrates that Peking University's mathematics community has a vibrant academic ecosystem," noted Academician Gang Tian, BICMR Director. 70

The Dual-Agent AI Framework: Rethlas and Archon

The core innovation is a dual-agent system powered by large language models (LLMs) like GPT-5.4:

• Rethlas: The informal reasoning agent uses Matlas—a semantic search engine over 13.6 million arXiv statements—to explore literature, propose strategies, and construct natural-language proofs via trial-and-error, mimicking a mathematician's process.
• Archon: The formal verification agent employs LeanSearch for Mathlib queries (Lean 4's theorem library) to translate proofs into code, detect gaps, and synthesize fixes autonomously.

Lean 4, a functional programming language and theorem prover, ensures machine-checked correctness. This end-to-end pipeline required minimal human input—only providing paywalled PDFs. 68 69

Photo by Kelsey He on Unsplash

Step-by-Step: AI's Path to Disproving the Conjecture

1. Rethlas reformulates the problem and searches Matlas, discovering Jensen's 2006 Corollary 2.4 on completions of local UFDs with trivial generic formal fibers—a cross-domain gem.
2. Constructs counterexample T = ℂ[[x, y, z]] / (x² - yz), a complete Cohen-Macaulay domain with non-principal height-one prime Q = (x, y)T.
3. Builds local UFD A with completion ˆA ≅ T; proves A weakly quasi-complete via Farley's criterion.
4. q = Q ∩ A is principal (q = (a)); A/aA ≅ T/aT not a domain, hence not weakly quasi-complete, so A not quasi-complete.

Rethlas iterated through failed plans before succeeding, self-verifying informally. 85

Formal Verification: A Monumental Lean 4 Achievement

Archon generated a 19,448-line Lean 4 codebase across 42 files in ~80 hours—10x faster than an expert. It formalized six external papers, filled gaps (e.g., transfinite recursion via well-founded ordinals), and bypassed missing Mathlib concepts using Kaplansky’s UFD criterion. The repo verifies cleanly via Lean kernel and Comparator, using only standard axioms. 84 69

This scale rivals person-months of work, open-sourced for global use.

Implications for Chinese Higher Education

At PKU, this validates BICMR's interdisciplinary model, attracting talent amid China's push for AI-math fusion. It positions Chinese universities as global leaders, reducing reliance on foreign tools via domestic LLMs. Ruochuan Liu remarked: "This has shown us a possible form of deep integration between mathematics and AI... cultivating a new generation of young researchers." 70

Global Impact on AI-Assisted Mathematical Research

Building on DeepMind's 2021 knot theory insights, PKU's framework excels in open problems, not competitions. Cross-domain retrieval via Matlas/LeanSearch unlocks scattered knowledge, promising faster theorem proving. Benchmarks show superiority over Aletheia on FirstProof tasks. For more, see the arXiv paper. 68

Photo by Hat Trick on Unsplash

Challenges, Innovations, and Future Outlook

Challenges included logical gaps and missing libraries; innovations like autonomous rerouting and 8,000+ daily LeanSearch calls address them. Future: AI-Newton for physics laws, deeper domestic AI integration. PKU's success inspires Chinese universities to invest in AI4Math, fostering actionable insights for research careers.

• Benefits: Accelerates discoveries, trains hybrid experts.
• Risks: Over-reliance on foreign LLMs; ethical verification.
• Solutions: Open-source tools, national collaborations.

Stakeholder Perspectives and PKU's Role

Experts praise the paradigm shift. Gang Tian: significant demonstration value. Liang Xiao infused algebraic intuition. For Chinese higher ed, it signals self-reliance in frontier tech, with BICMR as a model for multi-field teams. Explore research positions at institutions like PKU.