Blueshift was elite in its niche: Minerva hit 50.3% on MATH when the prior published state of the art was 6.9%

Asked:

“How good was the Blueshift Team at Google Deep Mind?”

This report draws on 18 rows of academic and official evidence (Google Research, arXiv, NeurIPS proceedings, OpenReview) plus 46 broader context rows gathered on 2026-10-28. The clearest independently measurable result is Minerva (2022); claims about Blueshift’s role in Gemini and Gemma are firsthand but self-reported by its leaders, and no public team-level review or ranking of DeepMind teams exists.

Verdict: elite in its niche and materially influential on frontier-model reasoning — but no public team-level performance review, headcount-normalized productivity measure, or clean comparison against other DeepMind teams exists, so a precise organizational rank or numerical score is not defensible.

Measured dominance: Minerva versus the field on quantitative reasoning

Percent of problems answered correctly, as reported in the Minerva paper (Table 3). MATH uses majority voting (maj1@k); without voting Minerva 540B scores 33.6% on MATH. Blue = Minerva; grey = comparison models.

The evidence ladder, strongest to weakest

TIER 1

Measured

Minerva’s benchmark results above are published, peer-visible numbers from the original paper, presented at NeurIPS 2022. Minerva 62B also matched the 2021 Polish national math exam average (57%); the 540B model scored 65%.

TIER 2

Official

Google’s Blueshift team page highlights a portfolio beyond Minerva: BIG-bench (207 tasks, 400+ authors, 132 institutions), PaLM (540B parameters), Block-Recurrent Transformers, and Sharpness-Aware Minimization — SAM set a CIFAR-100 state of the art (10.3% error) at ICLR 2021.

TIER 3

Self-reported

Firsthand accounts by co-leader Behnam Neyshabur (CV, LinkedIn) say Blueshift was responsible for the reasoning capabilities of the first Gemini release, built a math-specialized Gemini 1.5 model, and improved reasoning in Gemini 1.5 Pro/Flash and Gemma 2. These product-attribution claims are not independently audited.

TIER 4

Caveats

The Minerva paper itself notes performance remained well below human level, answers could not be automatically verified, the model had no calculator or Python access, and correct answers could follow faulty reasoning. Many outputs were collaborations across Google Research, Brain, and DeepMind, so credit cannot always be isolated to Blueshift alone.

What the chart cannot say

Minerva’s false-positive rate is nontrivial but bounded: Minerva 62B produced under 8% false positives on MATH, with mistakes split roughly half calculation errors and half reasoning errors. research.google

Blueshift’s output ranged well beyond math models: curriculum-learning (25,000+ models trained), length generalization in LLMs, scaling-law estimators, and out-of-distribution failure theory all carry Blueshift affiliations at ICLR and NeurIPS. openreview.net

A co-leader’s self-reported record claims 40K+ citations and an h-index above 50, with four LLM papers presented at NeurIPS 2022 — strong individual signals, but not a team-level audit. neyshabur.net

Institutional bios describe team lead Adam Brown as a core Gemini contributor running Blueshift as DeepMind’s science-and-reasoning team — organizational context, not a performance measure. aspenphys.org

Evidence table: Blueshift-affiliated research

Contribution	Benchmark	Reported result	Key caveat	Source

Method: 18 evidence rows from academic and official sources (arXiv, Google Research, OpenReview, NeurIPS proceedings; 4 independent hosts) plus 46 broader context rows across many hosts, gathered 2026-10-28. Benchmark numbers are percent of problems answered correctly as reported in the cited papers; product-influence claims on Gemini/Gemma are self-reported by named team leaders. One unrelated WMT translation paper appears in the evidence set and is kept in the table for completeness. Long text fields are truncated for space; full detail at the linked sources. No team-level numerical score exists or is invented here.

This report was generated automatically by Keenable SELECT at a user's request, from publicly available web sources linked herein. Keenable does not review, verify, or endorse its contents and makes no representation as to accuracy, completeness, or timeliness; AI-based extraction may contain errors. Nothing in this report is investment, legal, financial, or other professional advice. All trademarks and referenced content remain the property of their respective owners; no affiliation or endorsement is implied. To report an error, rights concern, or request removal: legal@keenable.ai.