Google AI Helps Resolve Decade-Old Math Conjecture Through Multi-Agent Verification
Google deployed multi-agent teamwork and Gemini to prove the 2013 Courtade–Kumar conjecture, formalizing 50M Lean 4 lines alongside mathematicians.

On November 13, 2026, Google Research announced that its multi-agent systems, orchestrated through Antigravity's /teamwork feature and powered by Gemini 4 Argon, had participated in proving a central open problem in information theory that resisted resolution for more than a decade. The announcement by Google Fellow Vahab Mirrokni highlighted the complete formal verification of the 2013 Courtade–Kumar conjecture in Lean 4, resulting in approximately 50 million lines of code.
Yet the true significance of the breakthrough departs sharply from the familiar marketing narrative of an autonomous artificial intelligence independently solving advanced mathematics. The achievement reflects a disciplined engineering pipeline where specialized agents explored parameter spaces and falsified flawed intermediate hypotheses. Furthermore, machine intelligence generated numerical certificates alongside human mathematicians who directed the overarching theoretical strategy.
The Information Theory Puzzle Behind the 2013 Conjecture
Formulated in 2013 by Thomas Courtade and Gowtham Kumar, the "Most Informative Boolean Function Conjecture" asks which Boolean compression rule preserves the maximum mutual information when an n-bit sequence passes through a binary symmetric channel with crossover noise p.
Intuitively, one might suspect that a sophisticated aggregation rule—such as a majority vote across all bits—would better filter out channel noise. Courtade and Kumar conjectured precisely the opposite: the optimal strategy is simply a "dictator function" that blindly copies a single preselected coordinate and discards the rest, achieving an upper bound of $I(g(X);Y) \le 1 - H_2(p)$.
Proving that this simple coordinate projection remains unbeatable across all possible balanced and unbalanced Boolean functions presented severe analytical hurdles. For over a decade, mathematicians verified restricted cases but failed to establish a general bound. High-dimensional nonlinear couplings continually resisted traditional induction methods.
Multi-Agent Orchestration and the Long Proof Pattern
The collaborative framework deployed by Google, in partnership with researchers from the Chinese University of Hong Kong (CUHK), the University of Southern California (USC), and Carnegie Mellon University (CMU), operated far beyond conventional single-turn prompting. In its technical post, Google Research detailed how /teamwork organizes agents into a structured research workflow known as the Long Proof pattern. This architecture addresses two inherent flaws of frontier models: unpredictable generation variance and cognitive drift across long execution windows.
To combat context degradation, /teamwork strictly partitions context by construction, assigning each specialized subagent only the specific files and symbolic tools required for its discrete step. Candidate proposals from generative models are immediately routed to deterministic falsifiers—including Lean 4 type-checkers, numerical solvers, and test suites—while independent critic agents audit intermediate deductions without inheriting the prompt history or biases of the generator.
This adversarial mechanism proved critical early in the collaboration. Gemini models identified concrete counterexamples to an ambitious generalized conjecture previously published by the CUHK team, preventing researchers from pursuing an unworkable proof path. With that dead end eliminated, the human researchers formulated a candidate objective that reduced the core problem to a 14-variable nonlinear optimization bound, enabling Gemini 4 Argon to systematically derive analytical estimates and interval-arithmetic certificates across the parameter space.
Deconstructing the 50 Million Lines of Lean 4 Code
Public reception of Google's announcement fixated primarily on the dramatic metric of 50 million lines of formalized Lean 4 code. An exhaustive technical audit of the public repository (dpwoodru/general-courtade-kumar-lean) published by Sysnote provides essential clarity on what those files actually contain. The repository encompasses 45,510 Lean files totaling 50,373,106 lines, but human-readable proof code occupies only 1,359 files and 238,690 lines.
The remaining 44,141 files—encompassing 4.88 gigabytes of storage—consist entirely of machine-generated interval arithmetic certificates that verify numerical inequalities across partitioned domains using Lean's decide tactic. Rather than representing 50 million lines of creative mathematical deduction, the volume measures the sheer computational labor offloaded to the proof kernel to verify analytical bounds that would exhaust human peer reviewers.
Human trust in the overall proof does not require inspecting millions of auto-generated certificates. Instead, verification rests upon a concise 63-line statement in GeneralCK/Statement.lean. External auditors validated the entire tree against the Lean Formal Mathematics Reference Operator, confirming that 22,107,215 declarations passed external checking via con-leche between September 23 and October 15, 2026, establishing an unusually transparent reproducibility record.
Independent Provers and the Reality of AI in Mathematical Research
Placing Google's announcement in proper historical sequence further dismantles the notion of a proprietary monopoly on AI-assisted discovery. The formal scientific preprint, titled "A Proof of the Most Informative Boolean Function Conjecture" and co-authored by Zijie Chen, Amin Gohari, Adel Javanmard, Honghao Lin, Vahab Mirrokni, Chandra Nair, and David P. Woodruff, appeared on arXiv:2609.24931 on September 21, 2026—nearly two months prior to Google's product-focused post.
Crucially, September 21 witnessed another breakthrough. Dr. Vu Khac Ky of FPT University and Professor Tuan Tran of the University of Science and Technology of China independently published a distinct proof of the Courtade–Kumar conjecture. Dr. Ky subsequently clarified that while his team leveraged artificial intelligence to rapidly test hypotheses and eliminate unpromising directions, machine intelligence served as an exploratory filter rather than an authorial mind. The following day, researchers Hessam Mahdavifar and Ahmad Beirami posted a third independent solution, disclosing that their one-bit formulation was developed using generative models under close human guidance.
These concurrent milestones differentiate modern research tooling from earlier standalone experiments like Aletheia, which tackled closed contest problems in the FirstProof challenge using Gemini 3 Deep Think in February 2026. While Aletheia's self-verification loops informed the design of /teamwork, Courtade–Kumar demonstrates distributed adoption: global mathematicians across independent institutions are quietly weaving language models into their scratchpads to accelerate discovery.
Collaborative Verification as the Future of Scientific Computing
The successful closure of Courtade–Kumar signals a structural transition in computational science. For decades, computer-assisted proofs—beginning with the four-color theorem in 1976—relied on bespoke numerical programs that left traditional mathematicians uneasy about undiscovered compilation bugs or hardware glitches. Pairing neural agents with interactive proof assistants like Lean 4 resolves that epistemological tension by forcing every generative step through an unyielding logical kernel.
Future breakthroughs in complex optimization, statistical physics, and formal hardware verification will depend on this symbiosis. Machine intelligence provides the brute-force exploratory stamina to map high-dimensional landscapes and certify millions of sub-cases, while human researchers supply the theoretical abstractions that formulate solvable questions.
Written by Adriano Nascentes
Founder & Lead AI Systems Analyst · Verified Staff
Founder and editor-in-chief of Unknown Observer. Technology researcher and software engineer investigating foundation models, autonomous agent runtimes, cognitive architectures, and frontier AI safety.
Was this article helpful?
Your feedback helps our editorial team refine reporting depth and accuracy.



