Google DeepMind reports Gemini Deep Think assisting research across mathematics, physics and computer science

Google DeepMind reports Gemini Deep Think assisting research across mathematics, physics and computer science

Google DeepMind’s post is dated February 9, 2026. Thang Luong and Vahab Mirrokni describe Gemini Deep Think assisting expert mathematicians, physicists and computer scientists with research problems, from proposing mathematical results to finding counterexamples and explaining methods. Their account presents Aletheia, a math research agent, as a system that generates and revises candidate solutions, checks them with a natural language verifier, and can acknowledge failure. The authors also describe researchers using Gemini in collaborative work across algorithms, machine learning, economics and physics. Their examples include a counterexample to a conjecture and an explanation of why a machine-learning method works. The authors argue that these systems can act as scientific collaborators, with people guiding and verifying the work. They also distinguish levels of AI contribution: they say some results have been submitted to journals as “publishable quality,” while claiming no “Major Advance” or “Landmark Breakthrough.” Their account therefore makes a substantial but bounded case: AI may help researchers tackle difficult problems, while the authors’ own classification stops short of the highest significance levels. The evidence presented here does not establish how broadly these results will generalize or how much independent verification each requires.

Publication: 6f67d103-f34f-500a-0997-6363257285f2

attribution — pass: The post is correctly attributed to Google DeepMind and its bylined authors, Thang Luong and Vahab Mirrokni. Claims stay attributed to them throughout, for example 'the authors argue' and 'their account'. The page is a single article, not a listing.

Thang Luong and Vahab Mirrokni

summary — uncertain: The summary is mostly faithful and readable, but it describes the authors' case as AI working 'with people guiding and verifying the work.' That leaves out the source's explicit claim of fully autonomous results: an AI-generated paper made without human intervention, and autonomous solutions to four Erdős open questions. Smallest repair: add one sentence attributing those autonomy claims to the authors. The summary also never says plainly that Gemini Deep Think is a Google AI model mode, though readers can infer it.

A research paper (Feng26) generated by AI without any human intervention

date — uncertain: evidence.originalPublished is 2026-02-09, from a backfilled feed date, but the page dateline reads February 11, 2026. It is unclear which marks first publication. The reading discloses the discrepancy in its limitations, which is good. It adds no hindsight framing and presents no forecasts as outcomes. The source is a publisher page, so no archive disclosure is needed.

February 11, 2026 Research

support — pass: Each claim's excerpt appears verbatim in the source and supports the claim. The summary details are source-stated: the counterexample to a conjecture, the explanation of the ML method, the journal submissions and the absence of Level 3/4 claims. The question about generalization follows from the evidence and is not a leading claim. The page is not marked as edited after publication.

Gemini engineered a highly specific three-item combinatorial counterexample, rigorously proving the long-standing human intuition false.

media — pass: No media was selected, and no verified media candidates were supplied. Selecting none is appropriate.