Google's Gemini Flash solves three PhD-level math problems using a teamwork framework.

Without seeing the name, you’d surely think this was another elite model too powerful to be publicly released, showcasing an astonishing record:

Conducting research: Solving seven of the most challenging problems in mathematics and computer science in one go, submitting a 40-page proof that even the most rigorous machine verification could not find an error;

Engineering project: Built a highly realistic CPU simulator from scratch, successfully booting the system with an error rate of just 0.71%;

Wrote code: Optimized the core code of two major open-source libraries, Eigen and ParlayHash, with changes directly merged by the upstream maintainers.

This is the latest performance report unveiled by Google's Antigravity team on August 27.

In the Teamwork technical article, Google's Antigravity team unveiled three categories of achievements: mathematics, systems, and open source.

Surprisingly, the star of this show isn't a power-hungry model, but the small model focused on speed and affordability: Gemini 3.7 Flash.

Google officially stated: This is the first time a Flash-level model has produced doctoral-level mathematical research.

How can a budget model outperform higher-tier models?

The secret is not in the parameters, but in a multi-agent orchestration framework called Teamwork.

What Google truly wants to convey to the industry is: it’s not that Flash suddenly became smarter, but that the way work is organized has changed.

Who is the main subject of this report? Google’s in-depth technical article provides a very precise definition:

Seven results in mathematics and theoretical computer science, all initially achieved by Gemini 3.1 Pro in Teamwork's long-proof mode.

But the remarkable part is that three of these groundbreaking achievements were fully replicated by Gemini 3.7 Flash.

These three items are by no means peripheral issues added for padding: coreset construction for ℓp subspace approximation, dimension lower bounds for maximum inner product embedding, and Hadamard quantization that directly reduces the leading constant by approximately 5.93 times.

Each one is a legitimate open problem in academia.

The remaining four items are exclusively addressed by 3.1 Pro: the condition number lower bound for sparse convex optimization, the approximate optimal lower bound for prefix matrix factorization, Knuth's Cycles problem, and the Erdős unit distance problem reproduced independently under offline conditions.

Moreover, the TCSBench score of 71%, which set a new internal record at Google, was achieved by the powerful combination of 3.7 Flash and 3.1 Pro, surpassing the previous record of 67.7% set by 3.6 Flash paired with 3.1 Pro.

Among these, the signal truly worth our attention is:

As long as the framework is set up correctly, lightweight models like Flash can fully replicate the research achieved by flagship models.

This is enough to expand our understanding of what small models can achieve.

Teamwork is a multi-agent orchestration framework developed by the Antigravity team.

Just type /teamwork-preview, and Gemini will read the prompt, select the mode, and instantly assemble an “AI expert team” that can run for hours—or even days.

The mathematical results mentioned earlier all come from the Long Proof mode.

Its design philosophy is highly counterintuitive: instead of stacking parameters, it relies on a group of Flashes coming together to challenge, debate, and exploit each other’s weaknesses.

So how exactly do these AIs hold meetings? Breaking it down, there are four steps total:

Step 1: The frenzied competition of "strategy searching".

The system simultaneously incubates multiple candidate solutions, assigning each one a dedicated "devil's advocate" whose sole KPI is to refute the solution.

Interestingly, the heavily criticized proposal is never directly discarded; instead, it remains in the process, carrying all its opposing feedback with it.

After all, in a dead-end "wrong path," you often find the inspiration that can save you.

Step 2: Follow the map to precisely break it down.

Once a reliable strategy is selected, the system breaks it down into a set of interdependent subproblems and organizes them into a rigorous topology. Independent tasks proceed in parallel, while those with sequential dependencies follow in order.

Step 3: The cutthroat internal competition.

Within each sub-question, hold another elimination round: nodes read through the candidate proposals while reviewing sharp criticisms, collaboratively refining them into an upgraded version.

If the consolidation fails, rerun it with the accumulated objections until all vulnerabilities are fully addressed.

Step 4: “Cross-Round Learning” — Learn from your mistakes.

Failed drafts are left untouched for the next round; every major pitfall encountered by validators is meticulously recorded in the Trap Log.

All dead ends explored and conclusions reached are synchronized in real time to the shared knowledge base, available for everyone to access anytime.

Tournament network in Long Proof mode: Each candidate strategy is paired with a falsifier, and rejected paths remain in the process with their objections.

Going through this process, it’s less like a cold, impersonal superbrain and more like a meticulous academic meeting where no one can get away with cutting corners.

Here, every proposal must first endure several rounds of intense scrutiny—only the toughest, most resilient ideas will make it through.

This directly resolves the collective hysteria most traditional multi-agent systems are prone to:

In the past, if one AI accidentally threw off the rhythm, other AIs would blindly follow like echo chambers, building higher and higher on a faulty foundation.

The secret weapon of teamwork is turning "pointing fingers at each other" into a robust, inescapable system.

Of these seven achievements, the most striking and最容易被误读的 is Knuth's Cycles, proposed by Donald Knuth.

In fact, this question was already solved by AI earlier this spring.

Donald Knuth, 2023 Stanford Christmas Lecture.

At the end of February this year, Claude Opus 4.6 swiftly constructed an odd case in just about an hour, prompting Knuth to write two "Shock!"s at the beginning of his paper.

Subsequently, models such as GPT-5.3-Codex and GPT-5.4 Pro stepped in to complete the most challenging even cases.

By mid-April, Gartner explicitly confirmed in the revised paper that the even case was no longer in doubt.

What did Google do this time?

In simple terms, Google found two more elegant and simpler constructions for even cases, and simultaneously released the first lengthy proofs, spanning over 40 and 70 pages respectively.

That 40-plus-page rigorous proof has also passed Lean formal verification, leaving not a single flaw for machines to find.

This is certainly a substantial academic contribution, but its true significance lies in providing a more elegant proof, rather than truly breaking new ground from scratch.

This instead highlights the true strength of Teamwork:

It does not attempt to fix the "island intelligence" of any single model; instead, through institutionalized interaction and coordination, it completely overcomes the collaborative weaknesses of multi-agent systems—such as fragmentation and blind conformity—unlocking the power of "collective intelligence."

The same spotting mechanism—Google changed the approach and directly applied it to tackle hardcore engineering.

This technical deep dive explicitly states "using Gemini 3.7 Flash." The team built a cycle-accurate, out-of-order execution RISC-V CPU simulator from scratch.

Out-of-order execution is a standard feature in modern high-performance CPUs and also the most common cause of simulator crashes.

Teamwork proceeds in two steps: first, ensure the microarchitecture functions correctly by implementing an out-of-order pipeline and reorder buffer, successfully booting the xv6 operating system to the shell; then, align timing cycle by cycle.

The process of Teamwork's RISC-V simulator booting the xv6 kernel and entering the Shell.

The hardest hurdle, which Google calls the "silent execution gap."

The microarchitectural state of the simulator may drift subtly over hundreds of cycles, and by the time an architectural error is reported, the root cause is already impossible to trace.

The solution for teamwork is to sandbox the reference simulator Spike to prevent agents from cheating or plagiarizing, and to perform synchronized simulation with reconciliation at every step.

Ultimately, this simulator successfully passed over 100 RISC-V standard benchmarks, achieving an average cycle error of just 0.71% compared to BOOM hardware on unseen test loads.

Google discloses: Cycle alignment comparison between the Teamwork simulator and BOOM hardware shows no average error exceeding 0.71% on test workloads.

However, it’s important to clarify that this is a software-level simulator, not an RTL chip design, let alone actual silicon fabrication.

Compared to mathematics and simulators, the last category of results may seem the least impressive, but it provides the most robust evidence.

Eigen is a highly popular high-performance linear algebra library in the C++ ecosystem.

Teamwork identified a suboptimal implementation of a row or column matrix-vector multiplication and directly implemented a fast SIMD path.

In the concurrent hash table ParlayHash, Teamwork incorporates optimization ideas from Swiss Table, doubling the initial insertion throughput with 64 threads, increasing single-threaded overall throughput by 1.5 times, and reducing memory usage per element by 25%.

These two changes are not self-congratulatory, closed-door score-chasing—they are real code that underwent rigorous open-source code review and was officially merged into the upstream branch by external human maintainers.

More noteworthy than the score-running is a statement at the end of the mathematical paper: the proof was first generated by Google’s internal Gemini agent system, then verified and edited by the authors.

Agents are responsible for wildly exploring endless drafts, while humans are responsible for signing off and final approval. This is the most realistic division of labor in current AI research.

Google itself clearly states: These issues would typically take top experts months to solve; Teamwork compresses the trial-and-error cycle, but the steering wheel and final approval remain in human hands.

In the second half of AI research, it’s no longer about who has the larger model parameters, but who has built a better AI team.

The cheaper and more disposable a model becomes, the more valuable human review and oversight become.

In the past, people were problem solvers. Now, people are problem creators and validators.

Gartner found a handwritten proof for Claude, and upon learning that someone had verified it using Lean, he said, "That's really good," because he was "getting more prone to errors lately."

Even proofs by the 88-year-old Turing Award winner must pass through the verifier—AI-generated proofs are no exception.

The more tasks the machine handles, the more it will take on. Someone must always be present to oversee the approval process.

Reference materials:

https://antigravity.google/blog/teamwork-when-ai-becomes-a-research-partner

https://www-cs-faculty.stanford.edu/~knuth/papers/claude-cycles.pdf

This article is from the WeChat public account "New Intelligence Yuan," authored by ASI Revelation, edited by Yuan Yu.