DeepMind's Gemini Deep Think is tackling PhD-level math and physics

AI has been great at acing math olympiads, but solving pre-packaged exam questions is a completely different game from doing frontier scientific research.

Real research is messy, the literature is vast, and advanced fields suffer from severe data scarcity—conditions that usually cause frontier LLMs to hallucinate wildly. Now, Google DeepMind wants to prove that its reasoning models can actually do the heavy lifting in theoretical science.

An AI that knows when to give up

DeepMind published two new papers detailing how an advanced version of Gemini Deep Think is tackling professional-level problems across pure math, theoretical physics, and computer science.

At the core of the mathematics effort is an agent codenamed Aletheia. To prevent hallucinations in specialized domains, DeepMind equipped it with a natural language verifier to catch errors in intermediate steps and iteratively refine candidate proofs. Crucially, the team programmed the agent to admit failure—saving human researchers countless hours of chasing dead ends.

Paired with web browsing to pull verified literature, Aletheia showed that test-time compute scaling holds up beyond student competitions and into PhD-level problems on their internal FutureMath Basic benchmark.

The results speak for themselves: Aletheia produced a research paper (Feng26) autonomously without human intervention to calculate structure constants in arithmetic geometry, evaluated 700 open questions on the Bloom's Erdős Conjectures database, and solved four of them completely on its own.

Smashing 10-year-old conjectures

DeepMind also teamed up with outside domain experts on 18 research bottlenecks across physics and computer science, relying on an "Advisor" model where humans guide the AI through iterative "Vibe-Proving" sessions.

To prevent confirmation bias, researchers used balanced prompting—forcing Gemini to simultaneously hunt for proofs and counterexamples. In one notable case in online submodular optimization, theorists had spent a decade trying to prove an intuitive conjecture from a 2015 paper. Gemini ended the debate by engineering a precise three-item counterexample, proving the human intuition completely wrong.

Elsewhere, when progress stalled on classic discrete algorithms like Max-Cut and Steiner Tree, Gemini bridged unrelated fields, pulling continuous math tools like the Kirszbraun Theorem and measure theory to crack the deadlocks. In physics, it simplified complex singular integrals in cosmic string theory into clean, closed-form finite sums using Gegenbauer polynomials.

DeepMind notes that about half of these cross-disciplinary results are headed for top conferences—with one already accepted at ICLR 2026—while others are being prepared for journal submissions.

Why it matters

For theoretical researchers, Gemini Deep Think is shifting from an autocompleting assistant to a tireless research partner that can search vast mathematical spaces for edge cases and bridge disparate literatures.

We think DeepMind's playbook here is remarkably sharp. Instead of claiming artificial superintelligence in science, they proposed a formal taxonomy with the math community to grade AI contributions, openly stating that their current outputs hit Level 2 ("publishable quality") while making zero claims on Level 3 or Level 4 breakthrough discoveries. That intellectual honesty builds genuine credibility with academia.

The emerging bottleneck, however, will be human verification. Producing a complex proof is becoming computationally cheap, but peer-reviewing an AI's novel mathematical reasoning takes immense human effort. As tools like Aletheia become widespread, scientific communities will soon face a tidal wave of machine-generated proofs that push human referees to their limits.