1d:I[70616,["/_next/static/chunks/0kcf5dt6yxdox.js","/_next/static/chunks/1su5n1uv1ti4s.js","/_next/static/chunks/0gvke1wv6udfg.js","/_next/static/chunks/1dfrle3jgmlzh.js","/_next/static/chunks/162lgvduyr-fa.js","/_next/static/chunks/2gcnnopam8e9w.js","/_next/static/chunks/2g60iab5a1pkg.js"],"Breadcrumb"] 1e:I[90762,["/_next/static/chunks/0kcf5dt6yxdox.js","/_next/static/chunks/1su5n1uv1ti4s.js","/_next/static/chunks/0gvke1wv6udfg.js","/_next/static/chunks/1dfrle3jgmlzh.js","/_next/static/chunks/162lgvduyr-fa.js","/_next/static/chunks/2gcnnopam8e9w.js","/_next/static/chunks/2g60iab5a1pkg.js"],"ArticleGrid"] 1f:T15e4,

Google researchers have published Stellar Colosseum, a many-agent harness that allocates inference across mathematics and theoretical computer science research, reporting 71.0% accuracy on a benchmark of research-level theorem-proving tasks drawn from FOCS, STOC and SODA papers. The paper, from a team including Honghao Lin, David P. Woodruff and Vahab Mirrokni, describes a system that is model-agnostic by design and has already been folded into Google Antigravity.

The premise is a familiar failure mode. Language models produce plausible short proofs but degrade on long-horizon problems, where progress depends on a chain of uncertain, interdependent decisions and one bad early commitment poisons everything downstream. Colosseum's answer is to stop treating a proof as a single trajectory.

Key takeaways

Stellar Colosseum scores 71.0% on TCS-Bench, a set of research-level theorem-proving tasks pulled from FOCS, STOC and SODA papers, using Gemini 3.1 Pro paired with Gemini 3.7 Flash.
In a separate Codeforces evaluation with Gemini 3.1 Pro, the proof-oriented pipeline with execution feedback solved 218 of 222 problems.
The workflow now ships inside Google Antigravity's Teamwork framework as the Long Proof pattern, available as a preview command on paid plans.

How the harness spends its compute

Colosseum runs several stages before it commits to writing a proof. It first explores alternative strategies rather than descending into the first one that looks promising, then applies a readiness gate that decides whether a given route is mature enough to decompose. Only then does it represent the proof plan as a set of interdependent section-level subproblems.

Verification is wired back into that structure. When a verifier flags a problem, the finding is routed to the specific part of the argument it affects instead of forcing a restart. Across every stage the harness generates candidates in parallel, attacks them with targeted falsification, and merges candidates and their critiques into a single research artifact through what the authors call overlapping random-sample tree aggregation.

That is a different bet from scaling a single agent's context and letting it think longer. It is closer to running a small adversarial research group, where the expensive resource being scheduled is inference rather than researcher time.

What the benchmark numbers say

TCS-Bench is a harder target than the contest-math sets that dominate model announcements, since its tasks are taken from papers accepted at three of the field's top theory venues. The 71.0% figure comes from a mixed configuration: Gemini 3.1 Pro alongside the cheaper Gemini 3.7 Flash, which suggests the harness recovers quality through coordination rather than by routing everything to the largest model.

The Codeforces result is narrower but striking — 218 of 222 problems solved with execution feedback in the loop. Both numbers come from the authors' own evaluation, and neither has been independently reproduced. The benchmark caveat that applies to every agent-scaffold paper applies here too: harness results are notoriously sensitive to prompt and tooling details that rarely survive reimplementation.

From paper to product

Unlike most research scaffolds, this one already has a shipping surface. Google says the Colosseum workflow is integrated into Antigravity's Teamwork framework as the Long Proof pattern, one of five patterns the system selects between, and that Teamwork has produced seven open-problem results including a sparse convex optimization question from JMLR and Knuth's Cycles Conjecture, verified in Lean.

The paper claims new results addressing open problems from FOCS and JMLR papers. That is a meaningfully different claim from benchmark accuracy, and it is the one the field will scrutinize hardest — as it did with OpenAI's 10,000-agent Navier-Stokes attempt, where the asterisks mattered more than the headline. Formal verification in Lean gives Google a stronger footing on at least part of that claim.

FAQ

What is TCS-Bench?

TCS-Bench is a benchmark of research-level theorem-proving tasks drawn from papers published at FOCS, STOC and SODA, three leading theoretical computer science venues. It targets open-ended research problems rather than competition exercises. Stellar Colosseum reports 71.0% accuracy on it using Gemini 3.1 Pro with Gemini 3.7 Flash.

Is Stellar Colosseum tied to Gemini?

No. The authors describe the harness as model-agnostic — it is a scheduling and aggregation layer over whatever model serves the calls. The published evaluations use Gemini 3.1 Pro and Gemini 3.7 Flash, so its reported numbers are specific to those models, not to the architecture itself.

Can developers use it today?

Indirectly. The Colosseum workflow is available inside Google Antigravity's Teamwork framework as the Long Proof pattern, offered as a preview command on paid plans. The paper itself is a preprint on arXiv and has not been peer reviewed.

a:["$","div",null,{"className":"mx-auto max-w-7xl px-4 py-8 sm:px-6","children":[["$","script",null,{"type":"application/ld+json","dangerouslySetInnerHTML":{"__html":"{\"@context\":\"https://schema.org\",\"@type\":\"BreadcrumbList\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https://ainewsway.com\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"#ai-theorem-proving\",\"item\":\"https://ainewsway.com/tag/ai-theorem-proving\"}]}"}}],["$","script",null,{"type":"application/ld+json","dangerouslySetInnerHTML":{"__html":"{\"@context\":\"https://schema.org\",\"@type\":\"ItemList\",\"@id\":\"https://ainewsway.com/tag/ai-theorem-proving#itemlist\",\"name\":\"#ai-theorem-proving Articles\",\"numberOfItems\":1,\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"url\":\"https://ainewsway.com/articles/googles-agent-swarm-hits-71-on-research-level-math-proofs-mu2gbvs1\",\"name\":\"Google's Agent Swarm Hits 71% on Research-Level Math Proofs\"}]}"}}],["$","$L1d",null,{"items":[{"label":"# ai-theorem-proving"}]}],["$","h1",null,{"className":"mb-8 text-3xl font-bold text-zinc-900 dark:text-zinc-100","children":["#","ai-theorem-proving"]}],[["$","$L1e",null,{"articles":[{"id":"3423dac5-aa4c-4d19-945c-f7bc222c9f72","slug":"googles-agent-swarm-hits-71-on-research-level-math-proofs-mu2gbvs1","title":"Google's Agent Swarm Hits 71% on Research-Level Math Proofs","subtitle":"Stellar Colosseum spends inference on competing proof strategies instead of one long chain, and already ships inside Antigravity as the Long Proof pattern.","content":"$1f","excerpt":"Google's Stellar Colosseum harness scores 71.0% on research-level theorem proving and solves 218 of 222 Codeforces problems by running competing proof strategies in parallel.","coverImage":"https://cdn.ainewsway.com/uploads/2026/09/e980ca19-125c-4908-8b17-042aec7575c6.webp","coverImageCaption":"Mathematical notation on a blackboard, the kind of research-level proof work Google's Stellar Colosseum harness targets.","authorId":"6773f67d-d7fa-4b19-8f38-384f10ac761b","status":"PUBLISHED","publishedAt":"2026-09-15T09:11:09.460Z","viewCount":0,"likeCount":0,"readingTime":4,"isFeatured":false,"isBreakingNews":false,"metaTitle":"Google Stellar Colosseum Scores 71% on TCS-Bench Proofs","metaDescription":"Google's Stellar Colosseum harness hits 71.0% on research-level theorem proving and 218 of 222 Codeforces problems, and ships in Antigravity Teamwork.","metaKeywords":["Stellar Colosseum","Google Antigravity","TCS-Bench","Gemini 3.1 Pro","multi-agent AI","AI theorem proving","AI math research"],"ogImage":null,"aiSummary":"Google researchers published Stellar Colosseum, a many-agent harness that allocates inference across mathematics and theoretical computer science research, scoring 71.0% on TCS-Bench using Gemini 3.1 Pro with Gemini 3.7 Flash. Instead of extending one proof trajectory, it explores competing strategies, gates when to decompose, and routes verifier findings back to the affected section. The workflow already ships inside Google Antigravity's Teamwork framework as the Long Proof pattern, signaling that agent scaffolding is moving from papers to products.","aiTags":[],"sentiment":null,"articleType":"NEWS","videoEmbed":null,"videoType":null,"seoScore":95,"isAutoGenerated":true,"isBackfill":false,"pipelineStage":null,"copyrightScore":null,"similarityScore":null,"matchedSourceId":null,"createdAt":"2026-09-15T09:10:40.803Z","updatedAt":"2026-09-15T12:13:13.448Z","contentUpdatedAt":"2026-09-15T12:13:13.448Z","deletedAt":null,"author":{"id":"6773f67d-d7fa-4b19-8f38-384f10ac761b","name":"Seung Jung","username":"Seung Jung","avatar":null},"categories":[{"id":"e2b5fffc-de32-42aa-86fe-c75f219d6b4b","name":"AI & Machine Learning","slug":"ai-ml"}],"tags":[{"id":"75a3bf99-8d8d-4109-a75c-238ead02890e","name":"Gemini","slug":"gemini"},{"id":"b4959a7c-29fe-47b5-a58d-c6f2b9fc343e","name":"Benchmark","slug":"benchmark"},{"id":"a9083e67-1546-4fbe-9789-4e61748b6f9c","name":"AI Research","slug":"ai-research"},{"id":"a1e84bbd-9fbe-4fca-a3ec-b953dc753981","name":"Google Antigravity","slug":"google-antigravity"},{"id":"0c7589f4-59b8-4ccf-ba30-098e1803902e","name":"Multi-Agent AI","slug":"multi-agent-ai"},{"id":"1f64b053-4e72-4642-b8fe-96e111d56f90","name":"AI Theorem Proving","slug":"ai-theorem-proving"}],"translations":[{"id":"d2bcad52-b513-4f53-970b-97362b62a918","language":"es","title":"El enjambre de agentes de Google alcanza el 71% en demostraciones matemáticas de nivel de investigación","subtitle":"Stellar Colosseum reparte la inferencia entre estrategias de demostración que compiten entre sí, en vez de estirar una sola cadena larga, y ya viene dentro de Antigravity como el patrón Long Proof.","excerpt":"El sistema Stellar Colosseum de Google obtiene un 71.0% en demostración de teoremas de nivel de investigación y resuelve 218 de 222 problemas de Codeforces ejecutando estrategias en paralelo.","metaTitle":"Stellar Colosseum de Google logra 71% en TCS-Bench","metaDescription":"Stellar Colosseum de Google alcanza el 71.0% en demostración de teoremas de investigación y 218 de 222 problemas de Codeforces, y ya viene en Antigravity."},{"id":"014a8e0a-7c40-4404-bb03-72c913f886fc","language":"ja","title":"Googleのエージェント群、研究水準の数学証明で71%を記録","subtitle":"Stellar Colosseumは長い思考の連鎖を1本伸ばす代わりに、競合する証明戦略へ推論資源を割り振ります。すでにAntigravityにLong Proofパターンとして搭載済みです。","excerpt":"GoogleのStellar Colosseumは競合する証明戦略を並列に走らせ、研究水準の定理証明で71.0%、Codeforcesの222問中218問を解きました。","metaTitle":"Google Stellar Colosseum、研究水準の証明で71%","metaDescription":"GoogleのStellar Colosseumが研究水準の定理証明で71.0%、Codeforces222問中218問を記録。Antigravity Teamworkにも搭載されています。"},{"id":"9a26c428-2867-404c-b046-af847775c63c","language":"ko","title":"구글 에이전트 군집, 연구급 수학 증명에서 71% 기록","subtitle":"Stellar Colosseum은 긴 사고 사슬 하나 대신 경쟁하는 증명 전략들에 추론 자원을 나눠 쓴다. 이미 Antigravity에 Long Proof 패턴으로 들어가 있다.","excerpt":"구글의 Stellar Colosseum 하니스가 경쟁하는 증명 전략을 병렬로 돌려 연구급 정리 증명에서 71.0%를 기록하고, Codeforces 문제 222개 중 218개를 풀었다.","metaTitle":"구글 Stellar Colosseum, 연구급 증명 71% 달성","metaDescription":"구글의 Stellar Colosseum 하니스가 연구급 정리 증명에서 71.0%, Codeforces 222문제 중 218문제를 기록했다. Antigravity Teamwork에 탑재됐다."},{"id":"dd75090b-4d93-48ac-a2dc-7756a5935f26","language":"zh","title":"Google 代理群在研究級數學證明拿下 71%","subtitle":"Stellar Colosseum 不把算力押在單一條長推理鏈上，而是分配給互相競爭的證明策略，並且已經以 Long Proof 模式內建於 Antigravity。","excerpt":"Google 的 Stellar Colosseum 框架以並行競爭的證明策略，在研究級定理證明拿下 71.0%，並解出 222 道 Codeforces 題目中的 218 道。","metaTitle":"Google Stellar Colosseum 研究級證明達 71%","metaDescription":"Google 的 Stellar Colosseum 框架在研究級定理證明拿下 71.0%，Codeforces 222 題解出 218 題，並已內建於 Antigravity Teamwork。"}],"celebrities":[{"id":"5c572908-2f5b-47b3-95e2-13ed3e298bf7","nameKo":"구글","nameEn":"Google","slug":"google","profileImage":null}]}]}],"$L20"]]}]

#ai-theorem-proving

El enjambre de agentes de Google alcanza el 71% en demostraciones matemáticas de nivel de investigación

Suscríbete al Boletín

#ai-theorem-proving

El enjambre de agentes de Google alcanza el 71% en demostraciones matemáticas de nivel de investigación