Most "AI solves math" stories in 2026 are really "AI gets a higher score on a math test" stories — a benchmark, a leaderboard, a percentage point. The Google Research preprint that quietly went up at the very end of September is a different kind of claim: a system called Cogentic, built on top of Gemini, produced new results on five problems that were genuinely unsolved — open research questions in theoretical computer science that no one, human or machine, had answered before. Independent domain experts then checked the proofs by hand.
Quick summary: Cogentic is a multi-agent orchestration system from Google Research, built on Gemini 3.1 Pro or an early version of Gemini 4 Argon, that produced novel, expert-verified results on five open problems spanning online learning, auction theory, and mechanism design. The paper, authored by Yang Cai, Vineet Gupta, Yanchen Jiang, Christopher Liaw, Aranyak Mehta, Grigoris Velegkas and Di Wang, was posted to arXiv on September 30 and revised October 1, 2026. Most problems took roughly 100 Gemini calls to resolve; the hardest took around 1,000.
What Cogentic actually does, in plain terms
A single prompt to a large language model can often produce a plausible-looking mathematical idea — but for genuinely hard, open problems, a one-shot answer from even a frontier model isn't enough, and it's usually wrong in some subtle way. Cogentic's answer is to stop treating proof-writing as a single conversation and treat it instead like a small research team with strict process. An orchestrator assigns a population of independent "prover" agents to distinct proof strategies in parallel. Their draft proofs then go through adversarial verification: one reviewer checks a draft in isolation, a second reviewer reads all of that round's drafts side by side specifically to catch mistakes multiple provers share, and both have to approve before anything is accepted — operating under the working assumption that every step is wrong until proven otherwise.
The system doesn't throw away failed attempts, either. A persistent record logs every attempt and the specific objection it failed on, and a separate auditor agent extracts any correct intermediate steps (lemmas) from otherwise-rejected proofs, re-verifies them independently, and adds them to a shared, growing ledger that later rounds can build on. That ledger is arguably the real innovation here — it's what lets the system make cumulative progress across many rounds instead of restarting from zero each time, closer to how a human research group accumulates partial results over weeks (Source: arXiv paper).
The five problems — and why "AI did it" isn't the end of the story
The paper reports new results across three fields: online learning (how an algorithm should make a sequence of decisions when it only learns the outcome after acting), auction theory (the mathematics of designing auctions that produce good outcomes even when bidders act strategically), and mechanism design (designing the rules of a system — an auction, a voting scheme, a matching market — so that people's self-interested behavior still leads to a good collective outcome). These are not high-school or even typical graduate-level exercises; they're the kind of problems that sit in active research papers and conference submissions, the sort a specialist might spend months on without resolution.
- Online learning: sequential decision-making problems where feedback arrives only after a choice is made, and the system has to balance exploring vs. exploiting an unknown environment.
- Auction theory: designing pricing and bidding rules so that an auction remains efficient or revenue-optimal even when participants try to game it.
- Mechanism design: the broader discipline of designing the rules of an economic or social system so that even strategic, self-interested behavior produces a good outcome.
- Domain-expert verification: each Cogentic result was independently checked by specialists in the relevant field, and the authors maintain a public ledger of the verified results rather than asking readers to simply trust the model's output.
This isn't the only AI-and-math story — and that's worth being precise about
It's easy to blur this together with a separate, higher-profile effort: Google DeepMind's AlphaProof Nexus and Aletheia systems, which are also built on Gemini (3.1 Pro and Gemini Deep Think respectively). AlphaProof Nexus alone has been credited with resolving nine of 353 unsolved Erdős problems, including two that had stumped mathematicians for 56 years, plus dozens of OEIS integer-sequence conjectures; Aletheia is a separately scoped system with its own, different Erdős tally, according to reporting on those efforts. Cogentic is a distinct system with a distinct focus: rather than working through a large catalog of pre-existing open conjectures, it's an orchestration framework purpose-built for pushing on genuinely new research questions in game theory and learning theory, with its own verification pipeline. Treating the two as interchangeable would overstate what either team has specifically claimed (Source: CryptoBriefing coverage of the separate DeepMind effort).
What this doesn't mean: that AI can now autonomously do research-level mathematics without humans. Every one of Cogentic's results required a human domain expert to verify it before anyone could call it correct, and the heaviest problems still needed roughly 1,000 model calls with a carefully engineered verification architecture around them — not a single prompt. The headline is real, but the infrastructure behind it is substantial.
01What is Cogentic?
Cogentic is a multi-agent system from Google Research, built on Gemini 3.1 Pro or an early version of Gemini 4 Argon, that produced new, independently verified results on five open research problems in online learning, auction theory and mechanism design. The paper was posted to arXiv on September 30, 2026 and revised October 1.
02Did Gemini really solve problems humans couldn't?
It produced new results on problems that were genuinely unsolved in the research literature at the time, yes — but every result went through independent verification by human domain experts before being accepted, and the hardest problems required roughly 1,000 orchestrated model calls with a dedicated verification pipeline, not a single answer from a chat prompt.
03Is this the same as DeepMind's Erdős-problem work?
No. That's a separate effort (AlphaProof Nexus and Aletheia) focused on resolving a catalog of existing unsolved Erdős conjectures and OEIS sequences. Cogentic is a different system aimed at new results in online learning, auction theory and mechanism design.
04Can I use something like Cogentic for my own math or research problems?
Not directly — Cogentic is a Google Research prototype, not a public product. But the underlying model, Gemini, is available through Google's own tools and through platforms like LumiChats, where you can use it for rigorous reasoning tasks, though you should still verify important results yourself rather than trusting a single AI response.
If this kind of frontier reasoning capability is relevant to your own research, coding, or analysis work, LumiChats gives you access to Gemini alongside Claude, GPT and 40+ other models in one place — useful for comparing how differently each one approaches a genuinely hard problem, rather than betting on just one lab's judgment.
